CO2301 Games Development 1 Week 23 Revision Lecture AI Revision

1
CO2301 - Games Development 1Week 23 - Revision
LectureAI Revision

• Gareth Bellaby

2
Characteristics of the exam questions

• Describe the topic, e.g. what is a game agent?,
  What is a FSM?
• Be prepared to give an example, e.g. give an
  example of a game agent, provide an example of a
  FSM.
• Provide an example of use. How would it be used?
  Why would it be used?

3
Prepare to draw a diagram

• Be prepared to draw a diagram for every topic.

Attacking
No hit points
Attacks
Idle
Start
Dying
4
Pathfinding

• What is pathfinding?
• 7 algorithms
• Crash and Turn
• Breadth-first algorithm
• Depth-first algorithm
• Hill-climbing algorithm
• Best-first algorithm
• Djikstra's algorithm
• A algorithm

5
Pathfinding

• Don't need to learn the algorithms off by heart.
• However, you must be aware of the fundamental
  characteristics of each of the search techniques.
• You must be able to explain and discuss the
  algorithms.
• Each of the search techniques can be compared to
  the others.
• What are the relative advantages and
  disadvantages of each of the searches?

6
Pathfinding

• Description The characteristics of the
  algorithm.
• Is the algorithm guaranteed to find a solution?
• If the algorithm guaranteed to find the optimal
  solution?
• Efficiency speed, size of tree.
• Heuristic blind or directed?
• Cost does the algorithm employ cost?
• Application Where would you use the algorithm?
  What are the advantages and disadvantages of the
  algorithm?

7
Pathfinding

• Simplest algorithm is "Crash and Turn" move
  towards objective, if crash into an obstacle turn
  and move around its perimeter until you can move
  towards the objective again.
• Simple to implement.
• Can be effective in simple situations
• Guaranteed to work so long as all objects are
  convex.
• Think about using in combination with other
  methods.

8
Breadth-first algorithm

• Expands all of the nodes, layer by layer.
• Exhaustive search.
• Guaranteed to find solution, if one exists.
• Guaranteed to find the optimal solution.
• Inefficient because every node searched. The
  problem of combinatorial explosion.
• Does not employ a heuristic.
• OK for small problems where size is manageable.
  Useful if you definitely want to search all of
  the nodes.

9
Depth-first algorithm

• Only one node expanded. Single path followed.
• Need to include backtracking, otherwise it is
  pretty useless.
• Not guaranteed to find a solution. (But variants
  such as using a limiting wall and employing
  backtracking are guaranteed to do so).
• Not guaranteed to find the optimal solution.
• Overcomes combinatorial explosion because a
  smaller tree is generated.
• Does not employ a heuristic.
• May get lucky and, on average, faster than
  breadth-first.
• Used for problems in which optimality not
  important.
• Can be used in conjunction with other searches,
  e.g. for opportunistic searches.

10
Hill-climbing algorithm

• Current node examined to find the best successor.
  Path follows best successor.
• Better than depth and breadth.
• Basic algorithm is not guaranteed to find a
  solution.
• Foothills.
• Plateaus.
• Ridges.
• However, it is guaranteed to find a solution if
  backtracking is used, together with the option to
  move to a worse successor if no other
  alternatives exist.
• Not guaranteed to find optimal solution.
• Uses a heuristic.
• Variants include randomness, jumps.
• Useful for climbing a hill! (Or equivalent).

11
Best-first algorithm

• All unexpanded nodes recorded. Best overall node
  chosen.
• Uses a heuristic.
• Better than depth, breadth and hill-climbing.
• Guaranteed to find a solution.
• Not guaranteed to find the optimal solution, e.g.
  on a map with terrain costs.
• Useful for maps with no costs or equivalent kinds
  of problems.
• Can be employed in a game with moving obstacles.

12
Djikstra's algorithm

• Uses cost.
• Choose the successor node with the lowest cost.
  Cost is accumulated along each route.
• Guaranteed to find a solution.
• Guaranteed to find the optimal solution.
• Very efficient for maps in which all of the costs
  can be pre-generated.
• For a game with a large map can be inefficient
  because potentially a large number of
  alternatives may be considered.
• Inefficient with moving obstacles since cost
  would have to be regenerated.

13
A algorithm

• Uses cost and heuristic.
• Choose the successor node with the lowest summed
  cost heuristic. Cost is accumulated along each
  route.
• Guaranteed to find a solution.
• Guaranteed to find the optimal solution.
• An efficient algorithm.
• A commonly used games algorithm. Compare with
  Djikstra's. If cost can be pre-generated and
  stored efficiently then Djikstra's will be the
  better choice, otherwise use A.

14
A algorithm

• Admissability, , i.e. the heuristic never
  overestimates the distance.
• Graceful decay of admissability. Be strict about
  admissability at the beginning of the search but
  as the search is assumed to get closer to the
  destination relax the admissability criterion.

15
Other considerations

• Level of detail.
• Representations grids, waypoints, regions,
  points-of-visibility.
• Path smoothing.
• Smoothing is implemented through the use of
  interpolation - calculating points in between our
  known points.
• Looked at a method of generating a curve called
  the Catmull-Rom spline.
• Generate tangents based on the position of the
  sample points.
• Re-acquisition of waypoints.
• Map design.

16
Game Agent
think
sense
act
memory
(optional)
17
Game Agents

• Autonomous of semi-autonomous.
• Perceives the world and acts upon the world.
• Agents are situated in the world.
• Agents interact with the world.
• Suits genres such as FPS, but also think about
  games such as Viva Pinata.

18
Game Agents - sensing

• Perceiving the world in a similar (or analogous)
  way to that of the player.
• Seeing.
• An efficient vision algorithm would only
  consider
• objects that the agent is interested in
• objects within the viewing distance of the agent
• objects within the viewing angle of the agent
• objects within the unobscured LOS of the agent
• Hearing.
• Similar considerations to seeing, but need to
  discuss sound propagation, ambient sound, etc.
• Communication with other agents.

19
Finite State Machine

• Finite State Machine is a machine which
  represents behaviour as
• states
• transitions between states
• actions. The actions can occur within a state, on
  entry to a state, on exit from a state.
• Need a start state and an end state.
• You need to consider FSMs as a diagramatic
  technique.
• I provided a bunch of examples. Make sure you
  understand them and write similar examples of
  your own.

20
Search Methods

• Applying search techniques to problem solving.
• A traditional approach to games such as Chess or
  Draughts.
• The example I gave was the Towers of Hanoi.
• Representation of the problem as a set of states.
• Use lists and a tree structure (same as
  pathfinding). Use rules to generate the new node.

21
Search Methods

• Combinatorial explosion the problem whereby
  rules in combination give rise to a large number
  of alternatives.
• Less useful with computer games.
• Chess is a 2-player game, with certain outcome
  and perfect information. Most computer games are
  n-player, uncertain outcome and imperfect
  information.

22
Production Systems

• Rule-based system if ... then rules.
• Can backward or forward chain the rules.
• Procedural knowledge is operational, i.e. what to
  do when.
• Allows reasoning. Deduction is the process of
  reasoning from premises to conclusions. Remember
  being able to answer the question "Is Gerald a
  Giraffe".
• Used in games such as RTS, e.g. Age of Empires,
  Rise of Nations, etc.
• This is an example of a routine rather than a
  game agent.
• Semantic nets as a variant. Bird example.
  Remember the connection with OO methods.

23
Game characteristics
24
Dealing with complexity

• Simplification.
• Representation. Use a less detailed
  representation.
• Level-of-detail, e.g. different sized grids for
  pathfinding, rule based system can have different
  levels.
• Task Management (Tank Example)
• Sub-divide the problem into tasks.
• Prioritise tasks.
• Assign tasks by priority to each of the actors.
• Hierarchical Decomposition (Command Hierarchy) A
  different representation for each level of
  analysis. The higher levels in the hierarchy can
  feed down to the lower levels. Mimics levels of
  command.

25
Targeting

• Attempting to hit a moving target. The first case
  is an infinite speed projectile.
• Infinite speed construct a LookAt function. You
  need to be able to write down how the function is
  constructed. You need to understand
• length of a vector
• normalise a vector
• dot product
• cross product
• the relevant axes

26
Targeting

• The second targeting case is a finite speed
  projectile.
• Simple leading. Set of simplifications.
• Deriving the relative movement of the target
  using projection of a vector.
• Be able to write down the maths.
• Why is it done human-like behaviour.

27
Topic List

• There's a list of words and topics I gave at the
  beginning of doing the AI.
• Week 7 "Topics to be learnt".
• Look at this material as well.