1. The Illusion of Intelligence
Introductions
The Integration of AI and Level Design in Halo
Chris Butcher
Jaime Griesemer

2. The Illusion of Intelligence
Target Audience
Designers, AI Programmers, Halo Players
Session Overview
Discussion of Halo’s AI Design Goals
Details of the AI Implementation
Description of the Level Creation Process
Demonstration of Halo’s Production Tools
Introductions
This session is not a full post mortem of Halo
we’re not exactly sure how it got billed that way, but we apologize if that was what you were expecting
This session is a discussion of Halo’s AI targeted at Designers and AI Programmers
It will also be useful to anyone looking for insight into how to get past Level 3 on the hardest difficultly level
Today we are going to talk about:
The Initial design goals for Halo’s AI
how they changed over the course of the project
Our AI implementation, including the technical constraints and important framework elements
as well as how the level creation process emerged from that implementation
And at the end of the talk we’ll give a demonstration of our production tools
and try to show how all these ideas translated into what the player experiences
Before we start I want to point out that the title of this talk is “The Illusion of Intelligence” and not “The Reality of Intelligence”
If you came here for tips about how to make the enemies in your game intelligent, we don’t have any for you
we don’t know how to do that, but it sounds really hard
What we do have, though, is tips on how to make the enemies in your game _appear_ to be intelligent
Chris and I have lots of experience with pretending to be intelligent

3. Meeting Player Expectations
Novel situations
Total interactivity
Significant challenge
Methods
Heavy scripting
Extended Interface
Omniscient and relentless enemies
Modern video games have a number of expectations placed on them by consumers
Players expect Novel situations
If Pac Man came out today reviewers would say it had no replay value
the ghost AI is repetitive and predictable
there’s no exploration, no story and the mission goals are always just ‘Eat everything’
One method of achieving Novel situations is through Heavy scripting – Medal of Honor
Positive aspect of this method is that the game feels like a movie
negative aspect is that it feels like a movie, not a game
players get the sense that everything is pre-ordained
this method is also time-consuming and technically demanding on level designers
Players also expect Total Interactivity
People are tired of “one man against the mindless evil horde”
They want every character to acknowledge them and react to them
the Player wants to be able to manipulate other characters
A common method to accomplish this is through an Extended interface – Unreal Tournament Bots
Unfortunately the Player is often forced to be a manager instead of a hero
the end result is that the Player often ignores this option
Players also expect a significant challenge
Gamers have become sophisticated
many of them have played games in your genre for years and they want a challenge
this is especially true of reviewers and early adopters, so it’s even more important
Some games provide this challenge through omniscient and relentless enemies – Quake III
This method feels like bullying
nobody wants to have a reflex competition with an electron
it requires precise tailoring to player’s reflexes (which can be difficult in solo campaign)
and it also becomes a barrier to the mass gamer because they can’t improve their talent level and there are no alternatives offered to achieve success
Please don’t take this slide to be criticisms of the games I mentioned, I intentionally only chose games I really enjoyed as examples it’s just an attempt to briefly explain why we decided not to use any of these methods

4. Where Design and Code Overlap
Parts of the AI are clearly Design
Mission Objectives
Scripting
Object placement
we aren’t going to talk about any of those parts
Parts of the AI are clearly Code
Firing Patterns
Pathfinding
Animation
we aren’t going to talk about any of those either

5. Where Design and Code Overlap
Design Responsibilities
3 minute scope
Racial personalities
Strategic purpose
Code Responsibilities
30 second scope
Intelligent decisions
Instant reactions
What we are going to talk about is Combat Behavior
this is where Design and Code overlap
Ideally you want to borrow the intelligence of the designers to make the AI seem smart
but you want to use the generalization of code to make the intelligence procedural (you can’t ship a designer on a cd)
In Halo, the Designers were responsible for anything that had a 3 minute scope
By 3 minute scope we mean any aspects of the AI that are established over a period of minutes, rather than seconds
An example of this would be the Racial personalities of our characters
Grunts are _always_ comical and cowardly
Elites are _always_ tough and aggressive
Providing the AI with a Strategic purpose
These are the goals that motivate the AI’s actions
that aren’t represented in code
Secure the landing zone
Protect the flag
Code responsible for 30 second scope, things that the player experiences on a moment to moment basis
Making Intelligent decisions
This rock is good cover
This tree is good cover
This enemy tank is not good cover
Instant reactions
This grenade is going to explode, I need to dive away
The Player is aiming a rocket launcher at me, I should hide

6. Design Goals Intelligible Interactive Unpredictable Individual Level
Imitating the Player’s capabilities
Transparent thought process
Racial personality
Group Level
Obvious strategic goals
Clear racial roles
Individual Level
???????????????????????????????????????????????????????
Limited capabilities
The Player assumes the AI has capabilities that are similar to his own
The AI should
Sees things the Player can see
Knows things the Player can know
Do things the Player can do
Elites shields recharge exactly like the Player’s shields
The Converse is even more important
should not do things the Player _can’t_ do
should not see things the Player _can’t_ see
There are no Psychic AI in Halo
There are no AI with perfect aim
There are invisible AI, but we let the Player try out invisibility as well
This is only a guide line
There are some AI that can jump higher than the Player
or run faster such as the Flood
Rule of thumb should be that you can give an AI an ability the Player doesn’t have as long as the ability is recognizable and the limits of that ability can be easily predicted
Transparent thought process
It should be obvious why the AI is doing what it is doing
One of the ways we chose to make it obvious was to limit the AI to basic cause and effect decisions
We also gave them custom animations and dialog for each decision
These animations are best if they communicate what the decision was
but even if it can’t be communicated its important to alert the Player that a decision was made
Another way to make the AI intelligible is to give individuals Racial personalities
In Halo, Grunts are cowards and all their behaviors derive from that
Hide when targeted
Panic when Elite is killed
Elites are aggressive
They rarely hide unless the player hides
Berserk charge on massive damage
Establishing racial personalities is helpful in two ways, it makes the design process easier
its easier to make a caricature than a complex portrait
it saves you from muddying the waters and creating confusing behaviors
Racial personalities also help the Player
it makes behaviors the Player has seen easier to remember
it makes new behaviors easier for them to identify and predict
keep in mind the Player’s first impressions are what establish these racial personalities
The first time you see an Elite he is a terrifying bad ass
The first time you see a grunt you easily kill him
The first time you see a hunter he is sent in as the last line of defense (and they are accompanied by foreboding music)
To make our AI intelligible on a Group Level
Strategic goals
Players are naively confident in your AI programmer
They will assume your AI knows they are supposed to be protecting the tractor beam power switch until you prove that they are just standing around
We wanted to work hard and let Players continue to believe the AI has a strategic objective
We mostly faked this, but it gives the Player the feeling that his enemies are opponents, working against him, rather than just disinterested obstacles
Clear racial roles
Consistent racial roles allow you to give the Player feedback on how a group of enemies is going to react
Grunts are the fodder so when the Player sees them he you can assume the Covenant aren’t prepared
Elites are the leaders so when the Player sees them he knows the Covenant are interested in the area and are going to put up a fight

7. Design Goals Intelligible Interactive Unpredictable Impressed Fooled
React to the Player
Surprise, Anger, Awe
Fooled
Limited knowledge
Predictable reactions
Thwarted
Breaking point
Flee in Terror, Berserk, Retreat, Defensive State
We know from experience that AI that doesn’t react to the player isn’t fun
so we assumed that the more ways the AI reacts to the Player the more fun it will be
But we didn’t just add efficient combat reactions
we wanted to add lifelike reactions that made the AI engaging
We wanted AI that could be impressed
The player wants to be important, so we create an entire universe and make him the center of it
Part of being the center of your own universe is being the biggest bad ass in the room
In Halo, everyone has a reaction to seeing the Player for the first time
Grunts are surprised
their first reaction is usually to take cover
Elites get angry
their first reaction is usually to shoot at you and make sure you can’t escape
Marines are just impressed
They stare at you and make comments about how great you are
“Wow, it’s him.”
“I didn’t think he’d be so tall”
An enemy that can be fooled
The player wants to be a ninja, he wants to toy with his enemies and humiliate them
Psychics cannot be fooled
but AI that uses its eyes and ears can be manipulated
therefore its important to limit the AI’s knowledge
Another key ability of ninjas is to predict what their enemies will do
We helped the Player predict what the AI would do by giving them lots of simple cause and effect reactions
If I shoot over here, the enemy will investigate over here, so I can ambush them
If I throw a grenade into that room, the enemy will run out of the room, so I can force them out of cover
We also helped the Player predict what the AI would do by giving the AI predictable strategies
For instance, they have a straightforward search pattern
they move forward at a constant rate
and they don’t double-back
allowing the Player to manipulate them
An enemy that could be Thwarted
The Player wants to be Arnold Schwarzenegger
he wants to dominate his enemies
he wants them to know he beat them
To accomplish this we included the concept of a breaking point
A breaking point is a distinct change in behavior that indicates the AI knows it is going to lose
For example, when a Grunt sees enough friends die or sees an Elite killed he flees in terror
When an Elite takes a large amount of damage he loses his temper and charges the Player
This happens on the group level as well, when a squad takes enough casualties it will retreat
it will try to find a defensive position
the squad members will stay in cover longer and fight defensively
This tells the Player that he has won

8. Design Goals Intelligible Interactive Unpredictable
Discarded: Randomness
Reactive AI
Unpredictable player
Unpredictable situations
Unpredictable reactions
Analog Reactions
Position
Timing
Unpredictable does not mean random
Random factors in behavior are rarely good
they are confusing because they cloud the issue
they make it difficult for the Player to analyze what is happening
Unpredictable means “not repetitive”
For reasons of intelligibilty and interactivity we talked about already, we want the AI to be predictable
Ideally this means that the player is the only unpredictable element
he is going to do slightly different things each time
We used this in Halo to trigger a cascade effect
Unpredictable player
puts the AI in unpredictable situations
which causes unpredictable reactions
and this leads to a unique experience
This causes a feedback loop because the unique experience leads to more unpredictable Player reactions and starts the cascade all over again
Analog reactions
Script driven encounters are very predictable because the inputs available to a script are very digital
Is the player in the room or not?
Is the AI dead or not?
In Halo the encounters are driven by the Code, which has much higher resolution of inputs
Exactly how far from the Player is each Actor and does it have line of sight?
Exactly how much health does each Actor have left and how far is he from cover?
This allows small changes in the situation to be amplified by the AI to yield large changes in behavior

9. Technical Constraints
20 – 25 Actors
2 – 4 Vehicles
About 15% of Xbox CPU
Two-player Cooperative
Support the Design Goals
Individual Knowledge
Emergent Behavior
Understandable
actor = AI character
we aim to simulate a single battle with around 25 actors total on all teams, with several vehicles added in as well.
this constraint mostly comes from the cost of character physics and animation and the fact that actors will be shooting off weapons all the time. so this is the maximum number we can have in a single battle. we make an exception for the flood infection forms, up to 100 of those.
Xbox: pentium III 733 MHz CPU.
15% isn’t our original estimate. originally we thought somewhere around 10-20%.
two players may be fighting completely different battles at once, double the AI load
three directions that we felt were really valuable to the game. each of these influenced the technical design in a number of ways.

10. Making the AI Interactive
Individual Knowledge Model
Discarded: Complete Model
‘Real’ Perception
No cheating
Vision, Hearing,
Touch, ESP
Selective Memory
Local objects
Crucial objects
Persistent State
Can be fooled
motivation: fun and understandable rather than smart and unbeatable
every actor has their own model of the world that describes their knowledge of its state. partial knowledge as well. this picture shows a marine trying to find a covenant elite warrior. this information comes from perceiving the world in a physically plausible way
no cheating! the AI system only has access to the actor’s knowledge model, not the actual state of characters in the world
four of the six senses – vision, hearing, touch and ESP (through scripting only, and for knowing where friends are).
originally we intended this to be a complete model with all actors tracking perception information for everyone in their visual range.
we only model the sections of the world that are relevant to combat – characters, vehicles, and dangerous events like grenades and weapon fire.
too complicated – O(n^2) for n=25. instead actors only track characters that are important to them, n around 3 for friends, 5 for enemies
nice outcome: every actor’s decisions depend on their own local situation which is easier for the player to understand
this gives human players the ability to reason about an actor’s knowledge, which lets them take advantage of the AI thought processes, so they can feel like the coolest space ninja commando that there ever was

11. Making the AI Intelligible
Communication of Intent
Discarded: Hidden States
Inform the Player
Language, Posture, Gesture
Focus of Attention
React to the Player
Dialogue
Animation
the smartest AI in the world is useless if the player has no way of knowing what it’s thinking
this means hidden state which influences an actor’s behavior is a real obstacle to fun AI. if players don’t understand why an AI made a decision then it seems much less smart, and much harder to learn about and predict.
an actor’s internal state is made visible in three ways: language (spoken dialogue, even aliens can still convey mood, e.g. “wort wort wort” [we did this by reversing english speech for the aliens]), different animation postures based on intention (sneaking, running, hiding), and specific animation triggers (gestures, canned dive/roll movements).
individual knowledge model let us model what objects are interesting to actors and control their facing, aiming and looking directions. lots of lifelike behavior really cheaply – turning to look at new characters, watching friends as they walk by, investigating dead bodies, spinning around to track the player if he starts shooting, etc
(give props to sigma, ken perlin)
we try to convey not just actors’ thoughts, but also what they think about the player’s actions. this makes players really feel like they’re an involved part of the game

12. Making the AI Unpredictable
Emergent Behavior
Discarded: ‘Fuzzy’
Emotion System
Cause-Effect Stimuli
Discovery
Weapon Fire
Damage, Death
Rich World Simulation
Unforced Group Behavior
motivation: predictable effects that combine and cascade to give fluid and deep gameplay
started out thinking about actors as collections of four fuzzy variables - fear, anger, defensiveness and surprise
this causes actions to be generated from thresholds and decay rates of hidden variables that are driven by multiple inputs
hard for human players to make a mental picture, so we removed these
most thought processes are driven by discrete events and stimuli, which is far more understandable
we still retain an analog measure of perceived threat level, defensiveness
it helps that halo is a very rich simulation. we don’t need to add complexity to the AI, because there are many game world elements that are designed to be self-contained and interact well together. things can happen that we never dreamed of (jackal shooting elite with charged bolt, warthog jump movie, elite swinging sword and breaking glass causing him to leap through window)
because the simulation is very flexible, group level just feeds down instructions in order to achieve coordination and leaves individuals to make their own decisions.
actors can always take time out to dive out of the way of a grenade, say hello to the player or whatever’s appropriate based on their situation.

13. AI Implementation
Now we’re going to move on from theory and talk about our implementation
In keeping with the theme of our talk I’ll explain a little bit about how we divided the responsibilities between design and code during implementation

14. AI Implementation Design Battle Flow
The designers focused on the large scale flow of battle
Here you can see a wave of marines charging a defensive line of Covenant
we’ll talk about these little squares later
for now its only important to notice how the Covenant are all grouped behind their defenses
and that the Marines are charging them as a group

15. AI Implementation Design Technical Battle Flow Actions and Responses
The programmers worked on the small scale actions and responses of the individual actors
Here you see the marines selecting targets to fire on
deciding which targets they can see
and which targets are in cover

16. Design Implementation
Difficulty Level
Battle Lines
Playtest Feedback
Lifespan
Smarter = Tougher
Tougher = Smarter
It’s should be obvious that in most cases Smarter AI = Tougher AI
What we found from Playtests was that the reverse Tougher AI = Smarter AI
It seems obvious in retrospect, but the tougher you make the enemies the smarter they seem to the player
In one playtest we had enemies that were weak
They died easily and didn’t do too much damage to the Player
36% of the Players thought the game was Too Easy
only 8% of the Players thought the AI were Very Intelligent
In the next playtest we increased the hit points of the enemies
and increased the damage they dealt
0% of the Players thought the game was too easy
43% of them thought the enemies were Very Intelligent
(Our playtest guys tell us that 43% is a very good rating
apparently people are reluctant to admit that they think a computer is smarter than them
Weak Enemy Playtest
Too hard
12%
About right
52%
Too easy
36%
Very Intelligent 8%
Somewhat Intelligent
72%
Not Intelligent
20%
Tough Enemy Playtest
Too hard 7%
About right
92%
Too easy 0%
Very Intelligent
43%
Somewhat Intelligent
57%
Not Intelligent 0%

17. Design Implementation
Difficulty Level
Battle Lines
Playtest Feedback
Lifespan
Smarter = Tougher
Tougher = Smarter
Consistent Challenge
Negative Reinforcement
Discourage boring tactics
Reward experimentation
Another way we used difficulty level to improve the AI was by providing a consistent challenge
If the game goes from easy to hard and back again
players attribute the easy parts to poor AI
and the hard parts to evil level designers
by keeping the difficulty level consistent we avoided these pitfalls
Negative Reinforcement
Don’t be afraid to kill the Player
he can tell how he messed up
he can tell how to change his tactics
we tried to accomplish these goals through the death cam
In Halo the Player was equipped with a recharging shield
this allowed him to take temporary damage
so we could show him that he was using the wrong tactic without killing him
One of the tactics we wanted to discourage was the Assault Rifle Rush
it wasn’t fun because the battles were ended too quickly
our AI gets more accurate the longer the Players stays out of cover
a long charge will almost always get you killed
Another tactic we discouraged was the Melee Attack Rush
Our AI isn’t very good at shooting at someone moving quickly very close to them
we gave some of the characters powerful melee attacks
If you are going to punish boring tactics you had better reward experimentation
So in Halo we rewarded stealthy, careful Players by letting them do instant kill melee attacks from behind
We rewarded Players that carefully selected their targets by making the Grunts flee when an Elite was killed
We also rewarded accurate Players by letting them get head shots
and we rewarded Players that planned ahead by making some weapons much better at certain enemies than others
In addition, for most encounters in Halo there is the easy way and the hard way
A frustrated Player that engages in a little exploration might find a hidden ramp leading to a catwalk above their enemies
or a player that simply watches an encounter before he attacks might get the chance to throw a grenade into a large group of guys at once

18. Design Implementation
Difficulty Level
Battle Lines
Playtest Feedback
Strategic Spaces
Interconnectivity
Killing Zone
Attacking/Defending States
Aggressive Territory
Retreat Conditions
Defensive Fortification
The designers were responsible for establishing battle lines, ???????????????????????
To do this we focused on making strategic spaces
Interconnectivity
It’s very important to make spaces that highlight the strengths of your AI
A battle in a corridor doesn’t involve much strategy
so there are no intelligent things for the AI to do
An interconnected space allows the AI to flank the Player, making them seem more intelligent
It also allows the Player to flank the AI, giving him a chance to watch them react to his tactics
One weakness of an interconnected space, however, is that it tends to be chaotic
So we also needed to establish a Killing Zone
an open area between two positions with good cover
flanking is still possible in an environment with a killing zone, but the battle is much more directed
the killing zone also allows multiple enemies to fight at once without it degenerating into chaos
Attacking/Defending States
Another way the designers controlled the flow of battle was through Attacking and Defending States
Each group of actors has an aggressive territory that it wants to occupy
it also has retreat conditions, such as a number of casualties or the position of the Player
the retreat conditions cause a transition to a defensive state where the Actors retreat to a more defensible location
By carefully collapsing an encounter from a large aggressive territory to a small defensive cluster
we were able to capture the feeling of a dynamic battle
I should mention the AI system we developed can get much more complicated,
especially when you layer several of these attacking/defending states on top of one another
but we rarely used that complexity
For a large majority of encounters, I’d say around 80%, we didn’t need to get any more complicated than this example

19. Design Implementation
Difficulty Level
Battle Lines
Playtest Feedback
Things to Avoid
Subtlety
Looking Broken
Insufficient Challenge
Things to Refine
Communication
Animations
Engagement Distances
The final tool we used when implementing the design side of the AI was Playtesting
For our playtests we bring in between 25 and 35 non-hardcore game players who have never played the game before
we put them in a lab and let them play for an hour
then we have them fill out a survey about their experience
we’ve found this gives us the most objective results and lets us act on the feedback with the confidence that it is accurate
As a brief aside:
If you aren’t doing playtests you should be
if you are doing playtests you should be doing more playtests earlier in your development cycle
you can’t have enough playtesting
We actually have the designers watch the playtests
because nothing forces you to confront reality faster than watching 30 people bang their heads against a level
until they hate the level, they hate the designers and they hate the game
In fact, during an early playtest we had more than half of the people ask if they could stop playing and go home
The desire to get people to stop wanting to quit was a major motivation for our early design process
The most useful data we got from our playtests was a list of things to avoid
At the top of that list is Subtlety
If it isn’t totally obvious, it’s too subtle
Even if you make something as obvious as you can possibly make it, half the people will miss it the first three times they see it
In Halo the Grunts run away when an Elite is killed
Initially nobody noticed so we had to keep adding clues to make it more obvious
By the time we shipped we had made it so not only does _every single_ Grunt run away _every single_ time an Elite is killed
but they all have an outrageously exaggerated panic run where they wave their hands above their heads
they scream in terror
and half the time one of them will say “Leader Dead, Run Away!”
I would still estimate that less than a third of our users made the connection
Another thing to avoid is letting your AI Look broken
When it comes to AI most Players are looking for any excuse to break their suspension of disbelief
they will see flaws even where there aren’t any and jump all over them
For Example: Early versions of Halo’s AI had a short pause between seeing you and recognizing that you were an enemy
this was to allow people to get the jump on the AI or retreat if they were getting in over their heads
Playtest feedback showed that Player’s assumed this Feature was a Bug and they all said it made the AI look dumb
We ended up completely removing the perception time on all the characters
because even though it was intentional people interpreted it as the AI being broken
Insufficient challenge
I talked about this earlier but it’s important enough to reiterate
Easy enemies are assumed to be dumb
Playtests are perfect for establishing how difficult the AI should be to make it look smart
The other useful feedback we got from our Playtests was a clear idea of what things we could refine
We could make sure that the thought processes we were trying to communicate where in fact getting across
We also learned which of our animations were being interpreted correctly and which ones were confusing to the Player
One of the most important things we were able to refine because of our playtests was our engagement distance
Nothing breaks our AI faster than engaging it from the wrong range
Sniping makes the AI look oblivious to danger
fighting too much from melee range makes the AI look robotic and awkward
We were able to use playtests to tweak our weapon ranges and activation ranges
to keep the Player at the optimal engagement range for our AI

20. Anatomy of an Actor World Interface Knowledge Model
Information flow restricted
Knowledge Model
Layered analysis
Distributed over time
Generates stimuli
Decision Logic selects from Actions
all communication between AI and world is tightly controlled
two main components of the actor are information and reaction
information gathering takes place in a layered knowledge model where information flows from raw sensory data into more and more processed forms
perceptual data -> memory -> spatial analysis -> situational analysis -> persistent emotions
decision logic is fed by stimuli about the world which are generated as we process information
current situation + stimulus -> decision logic selects an action to perform, which generates specific motion instructions
information gathering involves a lot of number crunching and can be distributed over time
decision logic needs to be quick, procedurally driven and to happen frequently (we ran complete decision code every tick)
next I’m going to expand on what the structure of the decision code is, what’s inside that box

21. Decision Logic Enemies cause alert Behaviors activated by stimuli
Innate combat cycle
Behaviors activated by stimuli
Charge, flee, seek cover
Throw grenade, enter vehicle, check dead body
Each race has a Black Box for action selection
Grunts flee easily
Elites seek cover if hurt
Jackals carry shields
a finite state machine where an actor has a single action that they are performing (fight, flee, charge, etc)
these actions have associated persistent information about their current goals and some very simple rules for completion
default transitions between actions come from an alert cycle based on an actor’s perception of enemies
this provides only very basic behavior of ‘see enemy, attack, pursue’
all further control is provided by a toolbox of behaviors which is layered on top of the basic decision making
behaviors each have trigger conditions based on the reception of certain stimuli during the information gathering phase
behaviors can override each other, modify the current action or change an actor’s emotional state
differentiation between races of actors comes in the form of custom ‘black box’ code that enables and disables behaviors based on environmental conditions
these rules are very simple and typically less than a page of code

22. Location, Location, Location
“This is my goal. Where should I be standing?”
Need a discrete answer to a continuous problem
the map is a continuous region but we need discrete points in order to reason about them
designers place points grouped where they want a squad to be located
gives the high-level ability to define broad areas
allows low-level control of combat spacing and cover-seeking
the designers control the movement of actors through a space solely by assigning them different sets of firing points to use
much of this happens automatically through attacking/defending and retreat conditions
for complicated movement, scripting is used
each action uses firing points differently but they all use the same points
cover tries to find somewhere out of the target’s line of sight
searching tries to find somewhere near the target’s last known location
casting rays for this calculation accounts for about 60% of all AI cost
the rest is mostly pathfinding and perceptual bookkeeping
Solution: Firing Points
Weighted and selected
line of sight
distance to target
proximity of cover
friends and enemies
vehicles, grenades, etc
Senses environment by multiple ray-casting

23. Combat Dialogue From decisions and stimuli Hundreds per second
hurt, death, saw enemy, throw grenade, seek cover
Hundreds per second
Priority, context, uniqueness, relevance
Select random dialogue type
Nearby characters can reply
Used for flavor only
57 events
166 dialogue types
12 speaking characters
5147 recorded lines
events are generated from basically everywhere in the AI system
there is a hierarchical ordering of priorities
some a priori knowledge of which are the rarest triggers, so they have a higher priority
try to pick actors which the player can see, and events to talk about which the player probably noticed
events can pick from multiple different dialogue possibilities, for example if a grunt in a stationary turret kills a marine
grunt says “kill you, human”
sarge says “keep it steady, marines, we’ll beat them” – custom for group of marines
a nearby marine screams “no! you covenant bastards!” – custom by killer’s race
another marine shouts “someone take that thing out!” – custom by damage type
dialogue is only for the player’s benefit, all AI coordination and communication happens inaudibly through other channels (but often causes dialogue events when it does happen)

24. Demonstration

25. In Conclusion... Design Goals Design Implementation
The Illusion of Intelligence
Combat Behavior is where Design and Code overlap
To sum up!
Most importantly, we were trying to create the Illusion of Intelligence, not actual intelligence
and that combat behavior is where Design and the Code overlap
Our Design Goals were
to make the AI Intelligible by
giving them similar capabilities to the Player
and making their thought processes transparent
to make them Interactive by
letting the Player impress them, fool them and thwart them
and to make them unpredictable by
using the Player as the focus of the battle
and giving the Actors a gradient of choices
Our designers focused on the flow of battle while our coders focused on the Actor’s action and responses
We used the Difficulty level to
let the AI live long enough to appear smart and
discourage tactics that made the AI appear dumb
We manipulated the battle lines in an encounter
by using strategic spaces to emphasize the AI’s strong points
and using Attacking and Defending states to make battles dynamic
and finally from our Playtests we learned that
Subtlety doesn’t work
and there are specific attributes of our AI that we could tune to make it appear smarter
Design Goals
Intelligible
Interactive
Unpredictable
Design Implementation
Difficulty Level
Battle Lines
Playtest Feedback

26. In Conclusion... Technical Goals Technical Implementation
Flexible Systems  Emergent Behavior
But make sure it stays fun and comprehensible!
Technical Goals
Communication of Intent
Individual Knowledge Model
Unpredictability
Technical Implementation
Actor Structure
Decision Logic
Firing Points
Context-based Dialog

27. Any Questions?