EmoBot: Incorporating Emotion into Quake 2 AI

1
EmoBot Incorporating Emotion into Quake 2 AI
Artificial Intelligence in Game
DevelopmentEastern Washington University, Fall
2003

• Tye Hooley
• Bart Hunking
• Mike Henry

2
Our Project

• The project had several goals we wanted to meet
• Modify the AI used by Quake 2s bots
• Incorporate a particular AI approach
• Make the Quake 2 AI more intelligent
  (hopefully!)
• Develop a tutorial based on this experience
• Demonstrate the modified Quake 2 game

3
Project Decisions

• Many possible approaches to building AI into a
  game
• Decision will often depend on the type of game
  and what algorithms or methods are most effective
  for solving problems in that game
• Quality of AI may sometimes need to be balanced
  with game performance (e.g., speed) theres a
  trade-off
• Often, finite state machines are used as simple
  solutions
• One possibility for some types of games is to
  give AI players emotions
• We saw that bots in Quake 2 could be given
  emotional states and perform actions specific to
  their current states

4
Quake 2 Artificial Emotion Engine

• Our decision create a simple artificial emotion
  engine for Quake 2s bots
• Why did we choose this?
• Relatively new approach to game AI
• Allows us to do some interesting experimentation
• Demonstrates uncertainty in reasoning via fuzzy
  logic
• Relevant to a game such as Quake 2
• We named our new bot AI EmoBot
• Some of the ideas for our engine were developed
  from Ian Wilsons paper, The Artificial Emotion
  Engine, Driving Emotional Behavior

5
Overview of the Engine (1)

• Here are some of the guidelines we used in
  creating the engine
• The emotional state of EmoBot should depend on
  sensory information available in Quake 2
  variables
• An emotional state (for the purposes of the
  engine) has two components fear and aggression
• EmoBot should exhibit some behavior as a set of
  actions, based on its current emotion

6
Overview of the Engine (2)
Fuzzy logic
Game variables
Fear
Emotion
Aggression
EmoBot action 1
Behavior
EmoBot action 2

7
A Brief Overview of Fuzzy Logic (1)

• Traditional set theory
• An item belongs to a set completely or not at
  all
• Example
• Few 0, 1, 2, 3, Many 4, 5,
• The number 2 belongs 100 to Few and 0 to Many
• Fuzzy set theory
• An item belongs to a fuzzy set with some degree
  of membership µ between 0 and 100 (i.e., in the
  interval 0, 1)
• A fuzzy set is defined by an associated
  membership function in some domain

8
A Brief Overview of Fuzzy Logic (2)

• Example
• 3 fuzzy sets cold, warm, hot in a temperature
  domain
• 45 has µ of .75 in cold, .25 in warm, and .00
  in hot
• Simple membership functions (as above) are
  preferred

cold warm hot
40 45 60
temperature, F
9
A Brief Overview of Fuzzy Logic (3)

• A fuzzy rule maps fuzzy sets from one or more
  input domains to a fuzzy set in an output domain
• A fuzzy rule set is a collection of these rules
• Example with two input domains
• input domain 1 temperature (cold, warm, hot)
• input domain 2 wind (light, moderate, strong)
• output domain time-outside (brief, long)
• IF temperature IS cold AND wind IS strong
• THEN time-outside IS brief
• IF temperature IS warm AND wind IS moderate
• THEN time-outside IS long
• etc.

10
A Brief Overview of Fuzzy Logic (4)

• Given a set of input values, the domains, and a
  fuzzy rule set, an output value may be found
• Step 1 Rule evaluation
• For each rule, memberships of the input values in
  the rules fuzzy input sets are found
• These determine the support for the rules fuzzy
  output set
• Step 2 Defuzzification
• Support results are combined and defuzzified into
  a single output value
• Example (cont.)
• Input values temperature 45F, wind 10 mph
• ? Rule evaluation ? Defuzzification ?
• Output value time-outside 90 minutes

11
Fuzzy Logic Process
Input domain 1
Input domain 2

Output domain
Input value 1
Input value 2
Rule evaluation

Fuzzy rule set
Defuzzification
Output value
12
Quake 2 and Fuzzy Logic

• Fuzzy logic for determining fear and aggression
  is implemented in Fril, which is an AI language
  useful for fuzzy reasoning
• Quake 2 was modified to make calls into the Fril
  code through a C-Fril interface
• Interface consists of a DLL and provides C calls
  into it

Quake 2 C code (modified)
Fuzzy logic Fril code
C-Fril interface
13
Division of C and Fril tasks

• Quake 2 C code Fuzzy logic Fril code

Quake 2 input values
Input domains
Aggr. output domain
Fear output domain
Behavior
Quake 2 AI actions
Rule evaluation
Fuzzy rule set
EmoBot emotion
Fear output value
Defuzzification
Emotion mapping
Aggression output value
14
Quake 2 Fuzzy Domains (1)

• 6 variables accessible in the Quake 2 code were
  used as input domains and each given 3 fuzzy
  sets
• bot health current health of the bot
• bad, fair, good
• player health current health of the player
• bad, fair, good
• bot damage amount of damage players weapon
  can inflict
• poor, average, great
• player damage amount of damage players
  weapon can inflict
• poor, average, great
• distance current distance from bot to player
• near, midrange, far
• angle angle between bots line-of-sight and
  line from bot to player
• small, medium, large

15
Quake 2 Fuzzy Domains (2)

• There are also two output domains with 3 fuzzy
  sets each
• fear level of bots fear
• low, medium, high
• aggression level of bots aggression
• low, medium, high
• For simplicity, all input domains and output
  domains use the same general model of membership
  functions
• All input values are scaled so that they lie in
  0, 1

0 .2 .4 .6 .8 1
scaled value
16
Quake 2 Fuzzy Rule Set (1)

• Fuzzy rules used to determine fear and
  aggression
• reasoning bot is not afraid when it has health
  to spare
• IF bot-health IS good
• THEN fear IS low-fear
• reasoning bots fear may increase if its
  health relative to player is lacking
• IF bot-health IS fair AND player-health IS bad
• THEN fear IS low-fear
• IF bot-health IS fair AND player-health IS fair
• THEN fear IS low-fear
• IF bot-health IS fair AND player-health IS good
• THEN fear IS medium-fear
• reasoning when bot is closer to death, fear
  becomes significant
• IF bot-health IS bad AND player-health IS bad
• THEN fear IS medium-fear
• IF bot-health IS bad AND player-health IS fair
• THEN fear IS high-fear

17
Quake 2 Fuzzy Rule Set (2)

• Fuzzy rules (cont.)
• reasoning bot should be aggressive with a nice
  weapon
• IF bot-damage IS great
• THEN aggression IS high-aggression
• reasoning bots aggression varies with a so-so
  weapon
• IF bot-damage IS average AND player-damage IS
  poor
• THEN aggression IS high-aggression
• IF bot-damage IS average AND player-damage IS
  average
• THEN aggression IS high-aggression
• IF bot-damage IS average AND player-damage IS
  great
• THEN aggression IS medium-aggression
• reasoning bots aggression varies more with a
  weak weapon
• IF bot-damage IS poor AND player-damage IS poor
• THEN aggression IS medium-aggression
• IF bot-damage IS poor AND player-damage IS
  average
• THEN aggression IS medium-aggression

18
Quake 2 Fuzzy Rule Set (3)

• Fuzzy rules (cont.)
• reasoning shorter proximity to player
  increases aggression
• IF distance IS near
• THEN aggression is high-aggression
• IF distance IS midrange
• THEN aggression is medium-aggression
• IF distance IS far
• THEN aggression is low-aggression
• reasoning longer time for bot to turn to
  attack increases fear
• IF angle IS small
• THEN fear is low-fear
• IF angle IS medium
• THEN fear is medium-fear
• IF angle IS large
• THEN fear is high-fear

19
EmoBot Emotion Space

• After the Fril code evaluates all rules and
  defuzzifies the outputs, two values are returned
• fear, in 0, 1
• aggression, in 0, 1
• This can be viewed as a point in two-dimensional
  fear-aggression space
• Regions of this space can then be associated with
  particular bot emotions
• We decided to divide this space into 6 regions
• The result is 6 separate emotions for EmoBot

20
Fear-Aggression Space

• Combination of 6 input values ? fuzzy logic ?
  fear-aggression x-y pair ? emotional state of
  EmoBot

21
EmoBot Emotional Behaviors (1)

• Each emotional behavior is composed of one or
  more Quake 2 AI atomic actions, which are
  performed through functions provided in the code
• Less aggressive behavior, low fear to high fear
• Cautious sniping attack
• Hold position (do nothing)
• Shoot
• Cautious Attack evasive attack
• Jump either left or right, chosen at random
• Shoot
• Afraid run away
• Turn 180 away from player
• Move forward

22
EmoBot Emotional Behaviors (2)

• More aggressive behavior, low fear to high fear
• Psycho confident rushing of player
• Charge player
• Shoot
• Cautious advance careful rushing of player
• Face player
• Move forward
• Shoot
• Strafe left or right, chosen at random
• Panic chaotic motion
• Turn bot toward random direction
• Move forward

23
End of Tutorial

• Specialized information provided after this slide

24
Weapon Damage (1)

• Much of the damage variable was based upon
  experience, although the basis was taken from the
  game. This means that we retrieved the values
  from the game and other references, then had to
  tweak our findings to make it more suitable for
  the EmoBots. Damage couldn't be based upon
  per-shot basis due to the fact that the guns have
  different rates of fire. So, damage/sec was
  calculated producing Table 1.1.
• We then used the calculated table as the basis
  for determining the actual EmoBot damage
  variables. These changes were based entirely
  upon game-playing experience. Table 1.2 is the
  final result of our tweaks.
• To get the final value between 0 and 1, we
  divided the weapon damage/sec by the largest
  weapon's damage/sec (BFG 250).

25
Weapon Damage (2)

• Table 1.1 Calculated Damage/sec
• Weapon Damage/sec
• -------------------------------------------------
  ---
• Blaster 40
• Shotgun 40
• Super Shotgun 120
• Machinegun 80
• Chaingun 120-240
• Grenade 41
• Grenade Launcher 120
• Rocket Launcher 96
• Hyperblaster 200
• Railgun 67
• BFG 166
• this is due to the lasers... they can do
  "unlimited" damage if the bot is around the
  projectile

26
Weapon Damage (3)

• Table 1.2 Our derived table--taking in
  consideration of player tactics
• Weapon Damage/sec
• -------------------------------------------------
  ---
• Blaster 40
• Shotgun 60
• Super Shotgun 100
• Machinegun 80
• Chaingun 110
• Grenade 70
• Grenade Launcher 120
• Rocket Launcher 150
• Hyperblaster 175
• Railgun 140
• BFG 250

27
Bot Special Attacks

• monster_berserk, 15 for spike attack,
• monster_gladiator, 20 for hit, 50 for railgun
• monster_gunner, same as gunner
• monster_infantry, 5 for hit, 3 for bullet
• monster_soldier_light, 2
• monster_soldier, 2
• monster_soldier_ss, 5
• monster_tank, 30 for blaster, 50 for rocket, 20
  for bullet
• monster_tank_commander, same as tank
• monster_medic, 2 for blaster
• monster_flipper, 5 for hit
• monster_chick, 10 for hit, 50 for rocket
• monster_parasite, 5 initial, 2 every iteration
  after
• monster_flyer, 5 for hits, 1 for each blaster
  shot
• monster_brain, 15 for hit
• monster_floater, 1 each blaster shot, 5 for hit
• monster_hover, 1 each blaster shot
• monster_mutant, 10 for hits
• monster_supertank, 50 for rockets, 6 per bullet

28
Damage by Bot Type, Scaled

• monster_berserk 0.50
• monster_gladiator 0.36
• monster_gunner 0.36
• monster_infantry 0.30
• monster_soldier_light 0.16
• monster_soldier 0.18
• monster_soldier_ss 0.20
• monster_tank 0.40
• monster_tank_commander 0.48
• monster_medic 0.16
• monster_flipper 0.20
• monster_chick 0.38
• monster_parasite 0.60
• monster_flyer 0.24
• monster_brain 0.75
• monster_floater 0.25
• monster_hover 0.17
• monster_mutant 0.32
• monster_supertank 0.60