Technical Issues: Artificial Intelligence

1
Technical IssuesArtificial Intelligence
2
Technical Issues Artificial Intelligence

• Artificial intelligence can mean a variety of
  different things in different contexts.
• By purist definitions, a game would possess
  artificial intelligence if a game player cannot
  distinguish between characters controlled by a
  human, or by the game itself.
• In such a case, the game would be passing a
  limited version of what is called the Turing
  test.
• In actual practice, however, a game that does not
  pass this test still has artificial intelligence,
  just not really good artificial intelligence.

3
Technical Issues Artificial Intelligence
Screen shot from Unreal Tournament 2003. If you
cannot tell if an opponent or teammate is a
human or a bot, then the bots artificialintellig
ence has passed the Turing test.
4
Technical Issues Artificial Intelligence

• Game developers rarely use the Turing test
  definition of artificial intelligence.
• In a game, artificial intelligence refers to the
  code used to control all non player characters
  and opponents within a game.
• The reactions of the game may be totally random,
  or totally logical, but the control code is still
  referred to as the artificial intelligence of the
  game.

5
Technical Issues Artificial Intelligence
Even though the opponent control for Centipede
(left) and block droppercode for Tetris (right)
is simple and scripted, with a random
number generator producing some variation, both
are still considered to posses a certain amount
of artificial intelligence.
6
Goals of Game Artificial Intelligence

• Players have different expectations of the
  artificial intelligence they find in different
  types of games.
• Some games, like Centipede and Tetris, have very
  minimal intelligence requirements.
• Other games, however, are heavily dependent on
  good artificial intelligence. If it is weak, the
  game would simply not be worth playing.
• Despite this, there are several general goals for
  any game artificial intelligence. The importance
  of these goals depends greatly on the game.

7
Goals of Game Artificial Intelligence
Screen shot from CompuChess. Without strong
artificial intelligence, a game of chess might
not be worth playing, except for beginners.
8
Goals of Game AIChallenge the Player

• Providing a reasonable challenge for the player
  must be the primary goal for artificial
  intelligence in any game.
• There are several ways to provide challenge
• A very sophisticated artificial intelligence.
• Outnumbering the player.
• Giving opponents abilities and advantages that
  the player does not have.
• Assigning the players teammates or additional
  obligations that might hold them back.
• Cheating. (As long as you dont get caught!)
• Poor game design. (Do not do this!!!)

9
Goals of Game AIChallenge the Player
Screen shot from Doom II. It created challenges
for players by vastlyoutnumbering the player,
and providing opponents many advantages
(unlimited ammunition, seeing in the dark,
flying, and so on).
10
Goals of Game AIChallenge the Player
Screen shot from Warcraft III. Sometimes, the
difficulty in selecting and controlling units in
the heat of battle provides an unwanted and
frustrating challenge.
11
Goals of Game AIChallenge the Player

• Creating a challenging and sophisticated
  artificial intelligence can be quite difficult.
• In some games, outnumbering the player and
  providing additional abilities is not what the
  player wants or expects.
• In such cases, the artificial intelligence must
  be very good.
• Depending on the game genre and game
  characteristics, the player must be challenged in
  different ways.

12
Goals of Game AIChallenge the Player
Screen shot from NHL 2003. It would break player
expectations by giving opponents extra abilities
or by outnumbering the player. The
artificial intelligence must be better to
compensate for this.
13
Goals of Game AIChallenge the Player
Screen shot from Alpha Centauri. Since it is a
strategy game, the player expects a strong
opponent as the game is very thought intensive.
Since thegame is turn based, the game cannot
overwhelm the player by processing alone the
player can take their time and think.
14
Goals of Game AIBe Realistic

• A game should provide artificial intelligence
  that is appropriate to its setting, story, and
  characters.
• Characters that are supposed to be smart should
  not do dumb things.
• Characters that are supposed to be dumb should
  not do smart things.
• The more human and realistic a character is, the
  smarter it should behave.
• Beast-like, alien, robotic, and undead characters
  can get away with more stupid actions, depending
  on the situation.

15
Goals of Game AIBe Realistic
Screen shot from Quake. The Zombies behave
pretty much as one would expect they lumber
towards you and take your shots until they get
close enough for an attack. Unless they are
blown to bits, they willget up and come back for
more, just like real zombies!
16
Goals of Game AIBe Realistic

• There are some things, however, that are so dumb
  that nothing should do it.
• For example, walking off of a cliff or not being
  able to navigate around a small obstacle.
• In these situations it is obvious to the player
  what the artificial intelligence should have
  done.
• Unfortunately, players seldom recognize how
  complex or difficult such obvious actions are to
  recognize and perform.
• To avoid ridicule, the games artificial
  intelligence must have a mastery of what is
  obvious to human players.

17
Goals of Game AIBe Realistic
Screen shot from Quake. The ogre was notorious
for getting stuck in doorways in many levels
with its chainsaw, and not knowing how to get
unstuck.To players, this seemed ridiculous, even
for an ogre.
18
Goals of Game AIBe Realistic
Screen shot from Oni. In this situation, Konoko
is being chased by Muro. A TCF officer on
Konokos side has beaten his enemy in the
background,and stands over her body for several
minutes. Why isnt he helping me?
19
Goals of Game AIBe Realistic

• One must be careful, however, to make artificial
  intelligence too realistic.
• Games are often unreal situations set up because
  they are interesting, not because they are
  authentic.
• For example, if an opponent realizes it has no
  chance of winning, it should run away
  indefinitely, which quickly ceases to be fun.
• In building good artificial intelligence, one
  must keep in mind the true goal of the project
  building a fun, playable game.

20
Goals of Game AIBe Realistic
Screen shot from 007 Nightfire. Realistically,
the villain should just killJames Bond and be
done with it, instead of toying around with
him. That wouldnt make for a very good game
though!
21
Goals of Game AIBe Unpredictable

• Humans are unpredictable. This is part of what
  makes them good opponents.
• The same should be true of the artificial
  intelligence opponents in a game.
• Players want the artificial intelligence to
  surprise them and use strategies and techniques
  that are unanticipated.
• If the player can predict with some measure of
  certainty what the game will do, the fun in the
  game quickly disappears.

22
Goals of Game AIBe Unpredictable
Screen shot from Starcraft. Strategy games
benefit greatly fromunpredictability. It does
not take long for a seasoned player to
recognize the same strategy over and over again.
23
Goals of Game AIBe Unpredictable

• Successful unpredictability can take many forms,
  depending on the game.
• Usually, this involves adding some element of
  randomness to the games artificial intelligence.
• Could be pure randomness.
• Could be a form of selection in which there are
  several valid choices of action that are chosen
  from randomly. Weights can be applied to vary
  the amount of randomness.
• In the end, the player will never know the action
  was random, and will tend to attribute it to some
  intelligence with a purpose.

24
Goals of Game AIBe Unpredictable
Screen shot from Unreal Tournament 2003.
Opponents can act inan unpredictable fashion
through randomly selecting a weapon to use,and
use tactics appropriate to that weapon.
25
Goals of Game AIBe Unpredictable

• Keep in mind that unpredictability should enhance
  the challenge presented by the artificial
  intelligence in a game.
• If things are so random that the game cannot put
  together a solid plan for defeating the player,
  you have gone too far.
• Make sure that random choices are still realistic
  given the scenario.
• If an opponent is about to win, and its
  artificial intelligence randomly selects a poor
  action, this will seem ridiculous.

26
Goals of Game AIAssist Storytelling

• The artificial intelligence in non player
  characters can be used for storytelling in many
  ways.
• Setting mood.
• Advancing the plot.
• Providing foreshadowing.
• And so on.
• By telling the story in the game, as opposed to
  just cut scenes, the player becomes much more
  involved.

27
Goals of Game AIAssist Storytelling
Screen shot from Oni. Konoko has just been
declared a rogue agent, and civilians encountered
reflect this to help establish setting and mood.
28
Goals of Game AICreate a Living World

• Creating a sterile game world filled with
  inanimate objects is not going to be a very
  authenticate reality for the player.
• Instead, adding ambient life to a world can do a
  lot to enhance the experience.
• Can be people going about doing their daily
  business.
• Can be birds flying in the sky, or animals or
  other critters roaming around the ground.

29
Goals of Game AICreate a Living World
Screen shot from Grand Theft Auto III. This
series of games always had avery lively city,
full of motorists and pedestrians with
interestingbehaviours. Too bad for them!
30
Goals of Game AICreate a Living World
Screen shot from the Legend of Zelda The Ocarina
of Time. The world is full of lively artificial
intelligence agents, including the characters
wandering around, animals, and so on.
31
Goals of Game AICreate a Living World
These artificial intelligence agents can be
interacted with in a variety of ways. This is
what happens if you attack poultry too often
death by chickens!!!
32
Artificial Intelligence Tips

• Before looking at some of the various techniques
  employed by artificial intelligence designers and
  coders, we first provide some general tips.
• Some of these are common sense.
• Others have come from years of experience, both
  good and bad.
• Unless you have a good reason, it is advisable to
  follow these suggestions!

33
Artificial Intelligence TipsDo Your Homework

• There is no one size fits all artificial
  intelligence system.
• Different techniques are appropriate for
  different situations.
• The right solution depends on many factors
• The emphasis of AI in the game.
• The scheduling and budget for AI development.
• The development team make-up and experience.
• It is important to determine the needs of your
  game and the capabilities of your team.

34
Artificial Intelligence TipsThe Sloped Playing
Field

• It is generally not advisable to pin all your
  hopes on creating an artificial intelligence that
  can compete with human players relying solely on
  its synthetic intellect alone.
• Generally, your artificial intelligence will
  require an edge to compete with a human player.
• By outnumbering the player.
• By having superior strength or abilities.
• By having additional resources at its disposal.
• By having perfect knowledge of the game world.
• And so on.

35
Artificial Intelligence TipsThe Sloped Playing
Field
Screen shot from Doom II. It makes use of a
sloped playing field by vastlyoutnumbering the
player, and providing opponents many advantages
(unlimited ammunition, seeing in the dark,
flying, and so on).
36
Artificial Intelligence TipsUse the Environment
Wisely

• Tune the design of levels to fit limitations of
  the artificial intelligence, instead of the other
  way around.
• It is incredibly difficult to build artificial
  intelligence that can handle every level design
  in an effective and efficient manner.
• Levels can be worked to fit with artificial
  intelligence limitations much easier and with few
  or no compromises.
• Your non player characters and levels have to
  work together to make a game fun to play.

37
Artificial Intelligence TipsUse the Environment
Wisely
Screen shot from Damage Incorporated. Players
had to direct teammatesthrough a 3D environment.
Levels had to be redesigned many times
beforerelease because of artificial intelligence
problems in navigation.
38
Artificial Intelligence TipsUse the Environment
Wisely

• Put smarts into the game environment itself, in
  addition to the characters.
• Have objects in the world announce themselves to
  non player characters and provide them
  information and scripts on how to use them.
• Embed navigation information in level design so
  that characters can traverse levels more easily.
• By making a smarter environment, the artificial
  intelligence driving characters does not need to
  be as sophisticated.
• Characters do not need to locate or identify
  objects or terrain they broadcast and tell them
  instead.

39
Artificial Intelligence TipsUse the Environment
Wisely
Screen shot from The Sims. The Sims makes use of
smart terrain thatbroadcasts what it offers.
For example, the refrigerator broadcasts that it
cansatisfy the hunger need and tells Sims how to
use it to do so when needed.
40
Artificial Intelligence TipsKeep it Simple

• In the rush to beef up the artificial
  intelligence in a game, it is very easy for it to
  get out of control and too complex.
• Ideally, the AI system should let the developers
  do complex things with a variety of simple parts.
• By combining the simple parts appropriately, a
  variety of interesting and complex behaviours
  can be provided.
• By keeping things simple, the system will be
  easier to understand, reuse, debug, and maintain.

41
Artificial Intelligence TipsPrecomputation is
Good

• Good artificial intelligence can be
  computationally expensive at run-time.
• Scarce resources are also needed for graphics,
  animation, physics, networking, and other
  subsystems though.
• Precomputation should be used wherever possible
  to provide good AI cheaply.
• Scripting of sequences of actions.
• Navigation through game terrain.
• Collisions with obstacles.

42
Artificial Intelligence TipsPrecomputation is
Good
Screen shot from Thief II. This game uses
navigation meshes to help characters navigate
terrain. By precomputing these in advance, and
usingthem in level design, character artificial
intelligence is simpler and cheaper.
43
Artificial Intelligence TipsTimeouts and
Fallbacks

• Nothing looks worse than an artificial
  intelligence that repeatedly does the wrong thing
  over and over.
• Players will not notice them make a wrong turn,
  but they will notice continuous collisions with
  an easy to navigate obstacle.
• Every artificial intelligence system should check
  for success conditions within a reasonable amount
  of time.
• If a timeout occurs, the system should give up
  and try something different.
• At a minimum, it can fall back to interesting
  idle animations that express its confusion or
  frustration while a new plan is formulated in the
  background.

44
Artificial Intelligence TipsTimeouts and
Fallbacks
Screen shot from Grand Theft Auto. Police were
notoriously bad at moving around stopped
vehicles to arrest the player they could easily
get stuck or run back and forth. Timing out and
falling back would have been good.
45
Artificial Intelligence TipsAvoid Story
Interference

• If the artificial intelligence in a game
  interferes with its story, it must be fixed.
• The artificial intelligence system must be aware
  of game events that are important to telling the
  story.
• This includes conversations, listening to dialog,
  watching a cut-scene, and solving game puzzles.
• The AI system should know it should back off and
  not get in the way.

46
Artificial Intelligence TipsAvoid Story
Interference
Screen shot from Oni. Konoko was having a
conversation with the scientist inthe lab coat
when she was viciously interrupted.
Unfortunately, she missedthe rest of the story.
The guard was punished appropriately.
47
Artificial Intelligence TipsProvide Memories

• The artificial intelligence characters in a game
  should remember what has happened to them and
  others during a game.
• They can then change their behaviour and dialog
  accordingly.
• This gives the player a sense that they are
  living beings with thoughts and feelings.
• If the artificial intelligence in a game can
  learn and adapt from its memories of events, so
  much the better.

48
Artificial Intelligence TipsProvide Memories
Screen shot from Quake 3 Arena. Most game
characters and bots havememories. If you shoot
them and get on their bad side, they rememberit.
They will even keep grudges against each other
too!
49
Artificial Intelligence TipsVariety Through Data

• A variety of character behaviours keeps games
  interesting and entertaining.
• Providing code for each behaviour introduces
  programming, debugging, and testing headaches.
• Instead, code should provide one or a small
  handful of behaviours that are greatly
  customizable through data.
• Designers can then introduce a new behaviour by
  tuning these variables.
• This includes awareness, speed, tactics, weapon
  preference, field and range of view, inventory,
  strength, abilities, chatter, and so on.
• This data should be available to designers to
  assist in game balancing and adjustments.

50
Artificial Intelligence TipsVariety Through Data
Screen shot from Unreal Tournament. It provides
a wide variety of bot behaviours based on the
settings of a few parameters.
51
Artificial Intelligence TipsSolve the Right
Problem

• Solving a different problem from the one thought
  to be at hand could result in a superior
  solution.
• Always ask yourself
• What is the real problem we are trying to solve?
• Can I solve a different problem that is easier or
  cheaper and get a result that is equivalent or
  superior?
• This process can be difficult but can also be
  very rewarding in the end.
• It takes time and practice to do this effectively.

52
Artificial Intelligence TipsSolve the Right
Problem
Screen shot from The Sims. To solve one problem
(How can Sims locate and identify game objects
to satisfy their needs?), Will Wright actually
solved the real problem (How do I get the Sims
to behave intelligently?).
53
Techniques for Artificial Intelligence in Games

• There are many different techniques used for
  providing artificial intelligencein games.
• Most techniques come from mainstream work in
  artificial intelligence.
• These techniques typically include
• Search techniques
• Rule based systems (finite state machines,
  decision trees, and production rule systems)
• Game theory and game trees

54
Techniques for Artificial Intelligence in Games

• Techniques used also include
• Artificial life and flocking techniques
• Planning techniques
• Fuzzy logic
• Scripting
• There are many other mainstream approaches that
  are not used in practice very much yet, but that
  is changing.
• This includes neural networks, genetic
  algorithms, belief networks, and so on.

55
Search Techniques

• Search techniques are used to serve a variety of
  functions in game artificial intelligence.
• Navigation and pathfinding.
• Constructing action plans (more later).
• And so on.
• These techniques are suitable for both
• Non player character control.
• Control of player characters not being directly
  controlled by the player themselves.

56
Search Techniques
Screen shot from Starcraft. Starcraft requires
navigation and pathfinding for directing both
player and non player characters around game
levels.
57
Search Techniques Navigation without Obstacles

• Without obstacles, navigation is a simple matter
  of determining which direction you need to go in
  to reach your objective.
• Some simple algebra will do the trick.
• If your movement directions are limited, pick the
  closest to the calculated result.

y
Start (x1,y1)
Goal (x2,y2)
x
Direction (x2-x1,y2-y1)
58
Search Techniques Navigation without Obstacles
Screen shot from Double Dragon. Navigation in
this case can be quite simple,as there are no
obstacles for non player characters to avoid.
They just needto home in on the player and
attack. (In later levels, there are
obstacles,but it is obvious that the AI did
little or no pathfinding around them!)
59
Search Techniques Navigation without Obstacles

• The same approach can also be used for aiming the
  shots of projectile weapons.
• With this however, you would also want to
  compensate for
• Motion of the target (velocity and acceleration)
• Motion of the shooter (velocity and acceleration)
• Motion of the projectile (acceleration, maximum
  velocity, friction, wind resistance, gravity, and
  so on).
• Depending on the skill of the shooter, you can
  tune how much compensation is actually done, and
  add random deviation to the chosen direction to
  avoid a perfect shot every time.

60
Search TechniquesNavigation with Obstacles

• This problem is significantly more complex than
  the case without obstacles.
• We now need our artificial intelligence to route
  characters around obstacles.
• The path selected must meet several criteria
• It must be as short as possible, or the least
  costly, depending on the situation.
• It must look realistic and make sense to select
  the path chosen.
• It must be selected in a reasonable amount of
  time without consuming too many scarce game
  resources.

61
Search TechniquesNavigation with Obstacles

• A leading approach to doing this is known as the
  A or heuristic search.
• There are several components to an A search
• A map (or graph) that A uses to find a path
  between two positions. Maps are composed of
  nodes that are connected if a direct path between
  the nodes exists.
• A cost associated with each direct connection
  between map nodes.
• A heuristic used to estimate the cost between two
  map nodes that are not directly connected. It
  must be an admissible heuristic that does not
  overestimate the cost.

62
Search TechniquesNavigation with Obstacles

• Each map node has three values associated with
  it
• g The cost to get from the start node to this
  node.
• h The estimated cost from this node to the goal
  node.
• f The sum of g and h representing our current
  best guess of the cost of the path going through
  this node.
• A maintains two lists of nodes an Open list,
  of nodes currently being explored, and a Closed
  list, of nodes that have been completely explored.

63
Search TechniquesNavigation with Obstacles

• The A algorithm
• Let S be the starting point.
• Compute g, h, and f for S.
• Add S to the empty Open list.
• Let B be the node from the Open list with the
  lowest f value.
• If B is the goal G, quit with success.
• If the Open list is empty, quit with failure.
• Let C be a node connected to B.
• Compute g, h, and f for C.
• If C is on the Open or Closed list, check if the
  new path is better with a lower f value, and
  update if necessary.
• Otherwise, add C to the Open list.
• Repeat 5. for all connections, then add B to the
  Closed list.
• Repeat from 4.

64
Search TechniquesNavigation with Obstacles

• There are many extensions to A
• Terrain based A that has variable costs based on
  different types of terrain.
• Hierarchical A to optimize paths for aesthetics.
• Tactical A to select paths based on certain
  strategies (avoiding enemy detection, to gain
  superior firing positions, and so on).
• Dynamic A that compensates for obstacle and goal
  movement throughout the map.
• There are also many other search algorithms
  besides A that are also effective.

65
Search TechniquesNavigation with Obstacles
66
Search TechniquesNavigation with Obstacles

• A can be expensive to compute, so we need to
  make some optimizations.
• The best way is to simplify the search space as
  much as possible.
• Use hierarchical pathfinding as much as possible,
  as it simplifies the search space.
• Make sure a good heuristic cost is used.
• Cautiously overestimate the heuristic cost, but
  you must be careful with this.
• Preallocate as many search resources as possible.
• Sometimes this is not the best approach

67
Search TechniquesNavigation with Obstacles

• An alternative approach is to use precomputed
  navigation meshes.
• An invisible mesh is used to represent all of the
  valid environment locations that can be occupied
  by game characters.
• This mesh can be traversed in an incredibly
  efficient fashion using a variety of algorithms
  to control character movement.
• Objects of use to characters can be placed along
  the mesh and searched for in an efficient fashion.

68
Search TechniquesNavigation with Obstacles

• Navigation meshes can be used to facilitate what
  would seem to be very complex and intelligent
  behaviour.
• Special points can be placed throughout the mesh,
  along with special links, indicating other
  actions can be successfully executed from those
  locations.
• This can be used to support jumping, crouching,
  crawling, diving, climbing, and a variety of
  other seemingly intricate behaviour with little
  or no additional effort.
• It makes characters seem smarter, but the smarts
  are part of the levels, not the artificial
  intelligence.

69
Search TechniquesNavigation with Obstacles
Screen shot from the Unreal Tournament 2003
editor. Paths and path nodesare added to ensure
that bots can navigate the environment. If they
areabsent, the bots will not be able to move
around and will get lost.
70
Rule Based SystemsFinite State Machines

• Finite state machines are simple and effective
  approach to game artificial intelligence.
• States represent the different game situations a
  character can face.
• On the occurrence of certain events, or with the
  passage of time, the character will take certain
  actions and will change to a new state. The new
  state can either be a different state or the same
  one.
• By encoding the appropriate events and actions in
  this fashion, state machines provide considerable
  power with little complexity.

71
Rule Based SystemsFinite State Machines
Opponent sighted in range Choose weapon, aim, and
fire
Not dead yet? Keep shooting
Running
Shooting
Opponent dead Go look for more
Out of ammunition Commence search
Hit jump point Start jumping
Done Go backto runsome more
Searching
Jumping
Simplified Finite State Machine for a First
Person Shooter AI
72
Rule Based SystemsFinite State Machines

• There are also several extensions to the finite
  state machine approach
• Hierarchical finite state machines allow states
  to be decomposed into their own separate finite
  state machines capable of implementing the parent
  state.
• Nondeterministic finite state machines allow
  multiple transitions on the same input event to
  different states. This can be used to add
  unpredictability to actions taken in a clean and
  consistent fashion. Weights can be added to tune
  the probability that the various transitions will
  be selected.
• Many more useful extensions exist.

73
Rule Based SystemsFinite State Machines
Been shot at Return fire
Running
Shooting
Been shot at Evasive maneuvers
Been shot at Start jumping frenzy
Dodging
Jumping
Nondeterministic Finite State Machine for a First
Person Shooter AI
74
Rule Based SystemsDecision Trees

• Decision trees are another approach to artificial
  intelligence.
• They are essentially complex if-then
  conditionals organized as a tree.
• Decisions based on a set of inputs are made by
  starting at the trees root and, at each node,
  selecting the next child node to visit based on
  the values of the inputs.
• The leafs of the tree contain the actions to take
  if tree traversal reaches the leaf.
• A variety of algorithms like ID3 and C4.5 can be
  used to construct such trees.

75
Rule Based SystemsDecision Trees
Am I being shot at?
Yes
No
Is it a gun or grenade?
Is an enemy in range?
Gun
Grenade
Yes
No
Aim and shoot!
Can one be found?
Strafe to dodge
Jump to dodge
Yes
No
Go after it
Get moreammunition
A Simplified Partial Decision Tree for a First
Person Shooter.
76
Rule Based SystemsProduction Systems

• Production systems are essentially a database of
  rules.
• Each rule consists of an arbitrarily complex
  conditional statement, and a set of actions to
  execute if the conditional is satisfied.
• Production systems are basically lists of
  if-then statements, with various conflict
  resolution mechanisms available in case more than
  one rule is satisfied at the same time.

77
Rule Based SystemsProduction Systems

• Some example rules for a production system for a
  First Person Shooter
• Enemy sighted
• Aim and shoot at enemy
• Out of ammunition
• Launch search for more ammunition
• Being shot at
• Execute evasive actions

78
Game Theory and Game Trees

• Some games can be thought of and treated as
  search problems quite effectively.
• The different possible game states are
  represented in a tree-like fashion. Different
  moves in the game trigger transitions from a
  parent state to its children states below.
• This tree is searched starting from its root (the
  current game state) to locate a game outcome
  resulting in a victory or the highest possible
  score for the player.
• The set of tree nodes traversed indicates the set
  of moves required to result in this situation.

79
Game Theory and Game Trees
Current Position
A1
A2
A3
3
2
4
Ply 1
A11
A12
A13
A21
A22
A23
A31
A32
A33
Ply 2
3
12
8
2
4
6
14
5
2
A sample game tree for a game. After white
explores two plies of moves ahead, its best move
would be A1, to which yellow would respond with
A11 for a valueof 3. If white made a different
move, it would end up with a value of 2.
80
Game Theory and Game Trees

• This approach can be very time consuming to
  explore and evaluate the various plies of
  possible moves.
• It is usually impossible with most games to fully
  explore all of the possibilities.
• Often, different approaches need to be used to
  improve overall efficiency.
• Take advantage of opponent think time to explore
  additional plies.
• Prune the tree whenever possible to reduce the
  number of possible moves to evaluate.
• Apply game-specific heuristics to be selective
  during the process of exploration. For example,
  in chess, spend extra time exploring
  queen-capturing moves.

81
Game Theory and Game Trees
Screen shot from CompuChess. Games like Chess
and Go frequently make use of game trees. They
can be applicable elsewhere too imagine a
fighting game that has planned moves and counter
moves several plies deep!
82
Artificial Life and Flocking Techniques

• Artificial life refers to systems of multiple
  artificial intelligence agents that attempt to
  apply some properties of living systems to
  themselves in a virtual world.
• For example, the agents can be given a variety of
  basic needs that they must fulfill. Fulfilling
  these needs typically requires other needs to be
  met, and often leads to other needs as well.
• Games such as SimCity and the Sims use these
  techniques to provide very interesting and
  entertaining gameplay.

83
Artificial Life and Flocking Techniques
Screen shot from The Sims. The Sims makes use of
artificial life techniques extensively. For
example, Sims feel hungry after a while and will
do what they can to satisfy that need, which can
lead to complex behaviours.
84
Artificial Life and Flocking Techniques

• Flocking is a subset of artificial life to
  regulate and direct group behaviour.
• Such systems employ the following general rules
  of flocking behaviour
• Separation steer to crowding local flockmates.
• Alignment steer toward the average heading of
  local flockmates.
• Cohesion steer to move toward the average
  position of local flockmates.
• Avoidance steer to avoid running into local
  obstacles or other terrain features.
• Survival steer to meet basic needs as necessary
  or to avoid danger if a predator seen.

85
Artificial Life and Flocking Techniques
Screen shots from several flocking Programs with
different behaviours.
86
Planning Techniques

• Planning techniques are an extension of search
  methods that emphasize the subproblem of finding
  the best sequence of actions to achieve a
  particular result.
• Start with an initial game world state.
• Also start with a precise definition of the
  consequences of each possible action.
• Planning then moves forward from the start, or
  backwards from the desired result to discover the
  sequence of actions required.
• Such techniques have met with some success.

87
Planning Techniques
Screen shot from Warcraft III. Game artificial
intelligence must formulate complex plans in
building units and structures, due to their
dependencies. (Other techniques could have been
used for the same result.)
88
Fuzzy Logic

• Fuzzy logic uses real-valued numbers to represent
  degrees of truth or set membership as opposed to
  the Boolean (true or false) values of traditional
  logic.
• Fuzzy logic techniques allow for more expressive
  reasoning and more richness and subtlety than
  traditional logic.
• They more accurately capture situations in which
  things are not either black or white.
• More often than not, things are shades of grey.

89
Fuzzy Logic
Screen shot from Grand Theft Auto. Fuzzy logic
is good for modeling trafficbehaviour with
multiple autonomous drivers in a game such as
this to control braking, steering, and
acceleration to avoid accidents.
90
Scripting

• Game artificial intelligence does not need to
  spontaneously think up every behaviour performed
  in a game.
• A combination of dynamic and scripted behaviours
  often works best.
• If the game designer can make assumptions as to
  what is happening in the game, scripted behaviour
  can be used quite effectively.
• For example, if a player is forced to enter a
  room from a certain direction, scripting can be
  used to provide an interesting challenge to the
  player.

91
Scripting

• When setting up scripts, the level designer must
  be involved.
• This way scripts can be written so that
  characters can take advantages of level features,
  such as hiding places, ambush locations, and game
  items.
• The game must also be able to detect when a
  scripted plan is not working out.
• It must switch to another script, or attempt some
  kind of unscripted action.
• Avoid predictability and ensure reactivity.
• Select scripts randomly, depending on the
  situation.
• Keep scripts short, and use them as building
  blocks.
• Try to build scripts that only provide hints of
  actions.

92
Scripting
Screen shot from Half-Life. This game was widely
praised for the strength of its artificial
intelligence. In fact, what was perceived as
intelligence was doneusing scripted paths that
game characters would follow in certain
situations.