Oct 6, Fall 2005Game Design1 AI in games Simple steering, Flocking Production rules FSMs, Probabilistic FSMs Path Planning, Search Unit Movement Formations

Oct 6, Fall 2005 Game Design 1

AI in games

Simple steering, FlockingProduction rulesFSMs, Probabilistic FSMsPath Planning, SearchUnit MovementFormations


AI in video games

5-10% of CPU for Realtime 25-50% of CPU for Turn-based

– Chase/Escape behaviors– Group behaviors– Finite State machines– Adaptation/Learning


Questions

How good should the AI be? Why are people more fun than

NPC’s? Will networked games reduce AI? New directions for AI in games?


Learning/Adaptation Increment aggressiveness if player

is doing well– The Computer-Based SATs do this!

Levels are a pre-programmed version of adaptation

Tuning Stability How might adaptation make play

Better or Worse?


Adaptation vs. Play quality

Do you want the monsters in Quake to get better as you get better?

Force the user to live with the consequences of his/her actions

Can surprise the designer (Creatures) Pit AI creatures against each other to

find bugs or tune actions Robotwar


What is good AI?

Perceived by user as challenging– Cruel, but fair!

User is surprised by the game– but later understands why

Feeling that reality will provide answers– able to make progress solving problem

What games have used AI effectively?


Chase/Evade

Algorithm for the predator?


Enhancements to Chase

Speed Control– Velocity, Acceleration max/min– Limited turning Radius

Randomness– Moves– Patterns


Enhancements to ChaseAnticipation

Build a model of user behavior


Steering Behaviors

Pursue Evade Wander Obstacle Avoidance Wall/Path following Queuing Combine behaviors with weights What could go wrong?


Group Behaviors

Lots of background characters to create a feeling of motion

Make area appear interesting, rich, populated


Flocking -- (HalfLife, Unreal)

What might go wrong?

Simple version:Compute trajectory to head towards centroid


Group Behaviors Reaction to neighbors -- Spring

Forces

Craig ReynoldsSIGGRAPH 1987


What could go wrong?

Repulsive springs around obstacles Does not handle changes in strategy

Exactly aligned Forces balance out in dead end


“Perceptual” Models


Production Rules

If( enemy in sight ) fireIf( big enemy in sight ) run awayIf( --- ) ----

Selecting among multiple valid rules– Priority weighting for rules or sensor events– Random selection

No state (in pure form)


Finite State Machines

States: Action to take– Chase– Random Walk– Patrol– Eat

Transitions:– Time– Events– Completion of action

Chopping

Take Wood to closest depot

Enough wood

Drop wood:

Go back to woods

At depot

At woods


State Machine Problems

Predictable– Sometimes a good thing– If not, use fuzzy or probabilistic state

machines Simplistic

– Use hierarchies of FSM’s (HalfLife)


Probabilistic State Machines Personalities

– Change probability that character will perform a given action under certain conditions

Aggressive PassiveAttack 50% 5%Evade 5% 60%Random 10% 10%Flock 20% 20%Pattern 15% 5%


Probabilistic Example

Fire At Enemy

Run out of Range

Enemy Within Hand-to-Hand Range

50%

Far Enough to Take Shot

Run Away

Enemy Within Hand-to-Hand Range

50%


Probabilistic State Machines

Other aspects:– Sight– Memory– Curiosity– Fear– Anger– Sadness– Sociability

Modify probabilities on the fly?


Planning

Part of intelligence is the ability to plan Move to a goal

– A Goal State Represent the world as a set of States

– Each configuration is a separate state Change state by applying Operators

– An Operator changes configuration from one state to another state


Path Planning

States:– Location of Agent/NPC in space– Discretized space

• Tiles in a tile-based game• Floor locations in 3D• Voxels

Operator– Move from one discrete location to next


Path Planning Algorithms

Must Search the state space to move NPC to goal state

Computational Issues:– Completeness

• Will it find an answer if one exists?

– Time complexity– Space complexity– Optimality

• Will it find the best solution


Search Strategies

Blind search– No domain knowledge. – Only goal state is known

Heuristic search– Domain knowledge represented by

heuristic rules – Heuristics drive low-level decisions


Breadth-First Search

Expand Root node– Expand all Root node’s children

• Expand all Root node’s grandchildren

Problem: Memory size

Root Root

Child1 Child2

Root

Child1 Child2

GChild1 GChild2 GChild3 GChild4


Uniform Cost Search

Modify Breadth-First by expanding cheapest nodes first

Minimize g(n) cost of path so far

Root

Child1 Child2

GChild19

GChild25

GChild33

GChild48


Depth First Search

Always expand the node that is deepest in the tree

Root

Child1

GChild1 GChild2

Root

Child1 Root

Child1

GChild1


Depth First Variants

Depth first with cutoff C– Don’t expand a node if the path to root

> C Iterative Deepening

– Start the cutoff C=1– Increment the cutoff after completing all

depth first probes for C


Iterative DeepeningRoot

Child1

Root

Child1

GChild1

Root

Child1 Child2

Root

Child1

GChild1 GChild2

Root

Child2

GChild3 GChild4


Bidirectional Search

Start 2 Trees– Start one at start point– Start one at goal point


Avoid Repeating States

Mark states you have seen before In path planning:

– Mark minimum distance to this node


Heuristic Search

Apply approximate knowledge– Distance measurements to goal– Cost estimates to goal

Use the estimate to steer exploration


Greedy Search

Expand the node that yields the minimum cost – Expand the node that is closest to

target– Depth first– Minimize the function h(n) the heuristic

cost function Not Complete! Local Minima


A* Search

Minimize sum of costs g(n) + h(n)

– Cost so far + heuristic to goal Guaranteed to work

– If h(n) does not overestimate cost Examples

– Euclidean distance


A* Search

Fails when there is no solution– Avoid searching the whole space– Do bi-directional search– Iterative Deepening


Coordinated Movement

Somewhat more difficult than moving just one NPC– Disappearing goal– New obstacles in path– Collisions with other NPCs– Groups of units– Units in formation


Coordinated Elements

Collision detection– Detection of immediate collisions– Near future

Perform the usual collision detection optimizations– Spatial hierarchies– Simplified tests– Unit approximations


Collision Detection

Levels of collision– Hard radius (small)

• Must not have 2 units overlap hard radius

– Soft radius (large)• Soft overlap not preferred, but acceptable


Collision Detection

With movement, need to avoid problems with bad temporal samples– Sample frequently

– Detect collisions with extruded units– Use a movement line– Detect distance from Line segment


Unit Line

Unit line follows path Can implement minimum turn radius Gives mechanism for position

prediction Connected line segments

– Time stamps per segment– Orientation per segment– Acceleration per segment


Prediction line

Given prediction, use next prediction as move

Prediction must have dealt with collisions already


Collision Avoidance Planning

Don’t search a new path at each collision

Adopt a Priority Structure– Higher priority items move– Lower priority items wait or get out of the

way Case-based reasoning to perform local

path reordering Pairwise comparison


Collision Resolution Summary

Favor:– High priority NPCs over Low Priority– Moving over non-moving

Lower Priority NPCs– Back out of the way– Stop to allow others to pass

General– Resolve all high-priority collisions first


Avoidance

Case 1: Both units standing– Lower priority unit does nothing itself– Higher unit

• Finds which unit will move• Tells that unit to resolve hard collision by

shortest move


Avoidance

Case 2: We’re not moving, other unit is moving– Non-moving unit stays immobile


Avoidance

Case 3: We’re moving, other is not:– If lower priority immobile unit can get

out of the way:• Lower unit gets out of way• Higher unit moves past lower to get to

collision free point

– Else If we can avoid other unit• Avoid it!


Avoidance

Case 3: We’re moving, other is not:– Else: Can higher unit push lower along

• Push!

– Else: Recompute paths!


Avoidance

Case 4: Both units moving– Lower unit does nothing– If hard collision inevitable and we are

high unit • Tell lower unit to pause

– Else: If we are high unit • Slow lower unit down and compute collision-

free path


Storage

Store predictions in a circular buffer If necessary, interpolate between

movement steps


Planning

Prediction implies – A plan for future moves

Once a collision has been resolved– Record the decision that was made– Base future movement plans on this

Blocking unit

Get-ToPoint

Predicted Position


Units, Groups, Formations

Unit– An individual moving NPC

Group– A collection of units

Formation– A group with position assignments per

group member


Groups

Groups stay together– All units move at same speed– All units follow the same general path– Units arrive at the same time

ObstructionGoal


Groups

Need a hierarchical movement system

Group structure– Manages its own priorities– Resolves its own collisions– Elects a commander that traces paths,

etc• Commander can be an explicit game

feature


Formations

Groups with unit layouts– Layouts designed in advance

Additional States– Forming– Formed– Broken

Only formed formations can move


Formations

Schedule arrival into position– Start at the middle and work outwards– Move one unit at a time into position– Pick the next unit with

• Least collisions• Least distance

– Formed units have highest priority• Forming units medium priority• Unformed units lowest


Formations

12

3

4

56

7

8

9

1 2 3

4 5

6 7 89

12

3

4

56

7

8

9

Not so good…

Better…


Formations: Wheeling

Only necessary for non-symmetric formations

1 2 3 4 5

1

2

3

4

5

Break formation hereStop motion temporarily

Set re-formation point here


Formations: Obstacles

1 2 3 4 5 Scale formation layout to fit through gaps

1 2 3 4 5 1 2 3 4 5

1 2 3 4 5Subdivide formation around small obstacles


Formations

Adopt a hierarchy of paths to simplify path-planning problems

High-level path considers only large obstacles– Perhaps at lower resolution– Solves problem of gross formation

movement– Paths around major terrain features


Formations

Low-level path– Detailed planning within each segment

of high-level path– Details of obstacle avoidance

Implement path hierarchy with path stack

High-Level Path

Low-Level Path1

High-Level Path High-Level Path

Low-Level Path2

High-Level Path

Low-Level Path2

Avoidance Path


Path Stack

High-Level Path

Low-Level Path1

High-Level Path High-Level Path

Low-Level Path2

High-Level Path

Low-Level Path2

Avoidance Path

1 2

Low-level path

Avoidance path


Compound Collisions

Solve collisions pairwise Start with highest priority pair

– Then, resolve the next “highest priority pair” now colliding


General

Optimize for 2D if possible Use high-level and low-level pathing Units will overlap! Understand the update loop

– It affects unit movement Maintain a brief collision history


References

Used with permission from CS4455 course at GA Tech by Chris Shaw

http://www.cc.gatech.edu/classes/AY2005/cs4455_fall/

Documents

Oct 6, Fall 2005Game Design1 AI in games Simple steering, Flocking Production rules FSMs, Probabilistic FSMs Path Planning, Search Unit Movement Formations