Maurice Bergsma and Peter Meij

1
Influence Maps

• Maurice Bergsma and Peter Meij
• Student Lecture GAI, May 3, 2006

2
Contents

• Influence Maps
• Reasoning with Influence Maps
• Sample Applications
• Recognizing Strategic Dispositions
• Environmental Awareness in Game Agents
• Issues with Influence Maps
• Conclusions

Introduction / Influence Maps / Reasoning / App 1
/ App 2 / Issues / Conclusions
3
Influence Maps

• Method for (runtime) spatial analysis
• Matrix representing the game world
• Each cell stores data for one aspect of that part
  of the world.
• Values propagate to nearby cells
• This shows the influence of certain objects on
  each part of the world
• Useful for many types of games

Introduction / Influence Maps / Reasoning / App 1
/ App 2 / Issues / Conclusions
4
Influence Maps (2)

• Different types of propagation are possible
• Linear
• Exponential
• Gaussian
• Etc.
• Game-specific, must be tuned
• Influences of similar objects can be added
  together

Introduction / Influence Maps / Reasoning / App 1
/ App 2 / Issues / Conclusions
5
Influence Maps (3)

• Occupancy

Introduction / Influence Maps / Reasoning / App 1
/ App 2 / Issues / Conclusions
6
Influence Maps (3)

• Occupancy
• Passability

Introduction / Influence Maps / Reasoning / App 1
/ App 2 / Issues / Conclusions
7
Influence Maps (4)

• Other possible types of influence maps
• Vulnerable assets
• Resources
• Exploration
• Line of fire
• Light level

Introduction / Influence Maps / Reasoning / App 1
/ App 2 / Issues / Conclusions
8
Reasoning with Influence Maps

• Multiple influence maps are stored in a spatial
  database
• These maps can be combined to create a
  desirability layer for a certain task
• A desirability layer contains the desirability to
  move to each cell

Introduction / Influence Maps / Reasoning / App 1
/ App 2 / Issues / Conclusions
9
Reasoning with Influence Maps (2)

• Methods of computing desirability
• Weighted Sum
• Simple and efficient
• Must be hand-tuned
• Opposing influences cancel each other out
• Product of normalized values
• Neural Networks
• Separate weights for each task
• Number of inputs is large for non-trivial tasks
• (layers x cells)

Introduction / Influence Maps / Reasoning / App 1
/ App 2 / Issues / Conclusions
10
Applications

• 2 Example applications
• Recognizing Strategic Dispositions
• Environmental Awareness in Game Agents

Introduction / Influence Maps / Reasoning / App 1
/ App 2 / Issues / Conclusions
11
Recognizing Strategic Dispositions
(image shamelessly copied from Nyree Viktor)
Introduction / Influence Maps / Reasoning / App 1
/ App 2 / Issues / Conclusions
12
Recognizing Strategic Dispositions
Introduction / Influence Maps / Reasoning / App 1
/ App 2 / Issues / Conclusions
13
Recognizing Strategic Dispositions
Introduction / Influence Maps / Reasoning / App 1
/ App 2 / Issues / Conclusions
14
Recognizing Strategic Dispositions
Introduction / Influence Maps / Reasoning / App 1
/ App 2 / Issues / Conclusions
15
Recognizing Strategic Dispositions
Introduction / Influence Maps / Reasoning / App 1
/ App 2 / Issues / Conclusions
16
Recognizing Strategic Dispositions
Introduction / Influence Maps / Reasoning / App 1
/ App 2 / Issues / Conclusions
17
Recognizing Strategic Dispositions
Introduction / Influence Maps / Reasoning / App 1
/ App 2 / Issues / Conclusions
18
Recognizing Strategic Dispositions
Introduction / Influence Maps / Reasoning / App 1
/ App 2 / Issues / Conclusions
19
Environmental Awareness in Game Agents

• Agents in games need situational awareness to
  give them a sense of
• What is happening in the environment
• What could happen next
• What options there are for action
• Possible outcomes of those actions
• Environment in the form of a grid
• Cells track important (threatening) events and
  conditions, such as rain, fire, other agents
• This can be done using cellular automata

Introduction / Influence Maps / Reasoning / App 1
/ App 2 / Issues / Conclusions
20
Environmental Awareness in Game Agents

• Cellular automata can be used in games to model
  simplified environmental processes, e.g. air,
  fluid flow
• Set of rules represent allowable physics of the
  model, and are used to update cell values
• New state of a cell is a function of previous
  state and state of neighboring cells

Introduction / Influence Maps / Reasoning / App 1
/ App 2 / Issues / Conclusions
21
Environmental Awareness in Game Agents

• Example automata

Introduction / Influence Maps / Reasoning / App 1
/ App 2 / Issues / Conclusions
22
Environmental Awareness in Game Agents

• Cellular automata in games
• Cells can include data for all kinds of game
  variables and are represented by continuous
  rather than finite values
• The rules and variables the cellular automaton
  requires depends on what is being modeled in the
  game environment

Introduction / Influence Maps / Reasoning / App 1
/ App 2 / Issues / Conclusions
23
Environmental Awareness in Game Agents

• Reacting sensibly to the environment requires
• Way to sense the environment
• Way to choose suitable reaction based on sensed
  data

Introduction / Influence Maps / Reasoning / App 1
/ App 2 / Issues / Conclusions
24
Environmental Awareness in Game Agents

• Role of influence maps
• Influence maps divide game map into a grid of
  cells, with multiple layers of cells containing
  information about the game world
• To include environmental decision-making one of
  these layers should consist of the data from the
  cellular automaton
• Also possible to use influence maps for tactical,
  low-level decision-making

Introduction / Influence Maps / Reasoning / App 1
/ App 2 / Issues / Conclusions
25
Environmental Awareness in Game Agents

• The reactive agents using the combination of
  influence maps and cellular automata can do this
  in three steps
• Compute danger function to determine each
  cells utility
• Use its return value to asses the level of
  reaction
• Let the agent choose the destination cell based
  on this

Introduction / Influence Maps / Reasoning / App 1
/ App 2 / Issues / Conclusions
26
Environmental Awareness in Game Agents

• Danger function

Introduction / Influence Maps / Reasoning / App 1
/ App 2 / Issues / Conclusions
27
Environmental Awareness in Game Agents

• Level of reaction

Introduction / Influence Maps / Reasoning / App 1
/ App 2 / Issues / Conclusions
28
Environmental Awareness in Game Agents

• Choosing a destination

Introduction / Influence Maps / Reasoning / App 1
/ App 2 / Issues / Conclusions
29
Environmental Awareness in Game Agents

• Agents are likely to have greater goals or
  desires they need
• This can be done through desirability values,
  provided through a desirability layer

Introduction / Influence Maps / Reasoning / App 1
/ App 2 / Issues / Conclusions
30
Environmental Awareness in Game Agents

• Propagation of desirability of a goal in an
  influence map

Introduction / Influence Maps / Reasoning / App 1
/ App 2 / Issues / Conclusions
31
Environmental Awareness in Game Agents

• Choosing a destination, also including
  desirability of the goal

Introduction / Influence Maps / Reasoning / App 1
/ App 2 / Issues / Conclusions
32
Environmental Awareness in Game Agents

• Algorithm for the agents
• Propagate desirability of the goal
• Each cycle, calculate the danger value for each
  cell in its local neighborhood
• Find the optimal cell in the local neighborhood
  of the agent
• Find the optimal cell in the immediate
  neighborhood

Introduction / Influence Maps / Reasoning / App 1
/ App 2 / Issues / Conclusions
33
Environmental Awareness in Game Agents

• Step 1 desirability propagation

Introduction / Influence Maps / Reasoning / App 1
/ App 2 / Issues / Conclusions
34
Environmental Awareness in Game Agents

• Step 2 Calculate danger value

Introduction / Influence Maps / Reasoning / App 1
/ App 2 / Issues / Conclusions
35
Environmental Awareness in Game Agents

• Step 3 Find optimal cell in local neighborhood

Introduction / Influence Maps / Reasoning / App 1
/ App 2 / Issues / Conclusions
36
Environmental Awareness in Game Agents

• Step 4 Find optimal cell in immediate
  neighborhood

Introduction / Influence Maps / Reasoning / App 1
/ App 2 / Issues / Conclusions
37
Issues with Influence Maps

• Time/space complexity
• Resolution of the grid
• Update frequency
• Tuning
• Parameter values
• Agent Design
• Overlapping influences
• 3D environments
• Suboptimal solutions

Introduction / Influence Maps / Reasoning / App 1
/ App 2 / Issues / Conclusions
38
Conclusions

• Influence maps provide an intuitive approach to
  (runtime) spatial analysis
• Influence maps are fairly easy and
  straightforward to implement
• Influence maps need quite a lot of tuning to
  deliver optimal solutions
• Influence maps are therefore not always
  guaranteed to be useful
• Influence maps can be applied effectively for
  strategic decision-making and reactive path
  finding

Introduction / Influence Maps / Reasoning / App 1
/ App 2 / Issues / Conclusions
39
References

• Sweetser, Penny. (2004) Strategic
  Decision-Making with Neural Networks and
  Influence Maps. AI Programming Wisdom 2, Charles
  River Media
• Sweetser, Penny. (2006) Environmental Awareness
  in Game Agents. AI Programming Wisdom 3, Charles
  River Media
• Tozour, Paul. (2001) Influence Mapping. Game
  Programming Gems 2, Charles River Media
• Tozour, Paul. (2004) Using a Spatial Database
  for Runtime Spatial Analysis. AI Programming
  Wisdom 2, Charles River Media
• Woodcock, Steven. (2002) Recognizing Strategic
  Dispositions Engaging the Enemy. AI Programming
  Wisdom, Charles River Media

40
Questions?