Using Cellular Automata and Influence Maps in Games Penny Sweetser The University of Queensland.

Using Cellular Automata and Influence Maps in Games

Penny Sweetser

The University of Queensland

Overview

Cellular Automata Influence Maps Grid-Based Techniques Decision making, environmental modelling Spread information in different ways Simple and powerful, separately or together Design, implementation, application to

games

Cells

Divide game world into cells Each cell a database containing info

about: combat strength, vulnerable assets, area

visibility, body count, resources, weather, passability

Cell size – accuracy / efficiency 10-20 standard units side by side

Influence Maps1

Strategic assessment / decision-making

Usually strategy games Spatial representation of AI’s

knowledge about the game world Strategic perspective of game state

layered over geographical1Tozour, P. (2001) Influence Mapping. In M. Deloura (Ed.), Game

Programming Gems 2. Hingham, MA: Charles River Media, Inc., pp. 287-297.

Influence Maps

Influence map indicates: where the AI’s forces are deployed where the enemy is / most likely to be where the “frontier” between players lies what areas are yet to be explored where significant battles have occurred where enemies are most likely to attack

in the future

Influence Maps

IM’s structure makes it possible to make intelligent inferences about: areas of high strategic control weak spots in an opponent’s defences prime “camping” locations strategically vulnerable areas choke points on the terrain other meaningful features that human players

would choose through intuition or practice

Influence Maps

IM tracks variables separately for each player (multiple parallel IM’s)

Each AI keeps one IM for itself and one for every other player

Could keep one IM and let all AI’s access it (but this is cheating)

Influence Propagation Once initial values given to cells, needs to

be propagated More accurate strategic perspective –

current influence / potential influence Spread influence with “falloff” rule Selection of falloff rules is subjective,

requires tweaking and tuning Exponential falloff – choose a constant 0..1 Need to terminate falloff (never reaches 0) Falloff should be proportional to cell size

Influence Propagation

2Sweetser, P. (2004) Strategic Decision-Making with Neural Networks and Influence Maps. In S. Rabin (Ed.), AI Game Programming Wisdom 2. Hingham, MA: Charles River Media, Inc., pp. 439-446.

Top-left: Game state

Top-right: Propagation

Lower-left: Influence values

Lower-right: Influence grey scale

Desirability Value

Estimates cell’s value with respect to a certain decision (e.g. where to attack)

Cells can be ranked by how good they appear for the decision

Usually calculated with weighted sum Choose relevant variables for decision Multiply by coefficient (roughly indicates

variable’s importance for decision) Sum all weighted variables together

Choice of variables / weights is subjective

Desirability Value

Variables used depends on game / design / decisions being made

Need to compensate for different units of measure (e.g. health vs. rate of fire)

Example desirability values: attack and defence desirability, exploration,

defensive asset placement, resource-collection asset placement, unit-producing asset placement, vulnerable asset placement

Weighted Sums for Desirability

Weighted sums are simple / transparent

But: Choosing the relevant variables is difficult Finding good weights is time-consuming Important info might be lost

Alternative to Weighted Sums

Simulated annealing or evolutionary approaches to find weights

Neural networks: Determine variables that most influence

decision / ignore irrelevant variables Variables are analysed in parallel, info in

individual variables is not lost Weights are determined during training

Neural Networks in IM’s2

Computational complexity Number of inputs and weights But don’t need to analyse whole map Train before shipping

Different AI personalities / strategies Learn to mimic human players

2Sweetser, P. (2004) Strategic Decision-Making with Neural Networks and Influence Maps. In S. Rabin (Ed.), AI Game Programming Wisdom 2. Hingham, MA: Charles River Media, Inc., pp. 439-446.

Cellular Automata in Games

Proposed as a solution to static environments in games3

More dynamic / realistic behaviour of scripted elements – fire, water, explosions, smoke, heat

Conducting research into using CA in games for environmental modelling

3Forsyth, T. (2002) Cellular Automata for Physical Modelling. In D. Treglia (Ed.), Game Programming Gems 3. Hingham, MA: Charles River Media, Inc.

Cellular Automata Research

No research or implementation of CA in games

Are CA appropriate for use in games? Can CA facilitate emergent gameplay? What effect will this have on the

player?

Cellular Automata - Traditional

Spatial, discrete time model Space represented as uniform grid Each cell has a state (from a finite set) Time advances in discrete steps Each step, cells change state

according to a set of rules New state = function of previous state

of the cell and state of neighbour cells

Cellular Automata - Traditional

1D – single line of cells, 2 neighbours

2D – 4 or 8 neighbours

1

2

1 2

Cellular Automata in Games

States are continuous (not discrete) E.g. heat = 657.21

States have multiple variables E.g. heat, pressure, water

Rules are continuous Damage = temp * burning rate

CA in Games Research4

Environmental systems Heat and Fire Rain and Fluid Flow Pressure and Explosions Integrated System

4 Sweetser, P. & Wiles, J. (unpublished) Using Cellular Automata to Facilitate Emergence in Game Environments. Submitted to the Journal of Game Development.

CA and IMs in Games

Cellular automata and influence maps can be integrated

Values generated by CA used for decision-making by influence map E.g. AI can consider environmental

factors when making a decision

CA & IM in Games Research

Agents used CA and IM to determine how to react to the environment

Agents use the cellular automata values to determine “comfort”

Added a goal (desirability) Desirability of goal is propagated

Conclusion

Grid-based techniques Cellular Automata Influence Maps

Advantages Allow type of behaviour to be specified

Disadvantages Lots of tuning / testing to get desired

behaviour

References

Forsyth, T. (2002) Cellular Automata for Physical Modelling. In D. Treglia (Ed.), Game Programming Gems 3. Hingham, MA: Charles River Media, Inc.

Sweetser, P. (2004) Strategic Decision-Making with Neural Networks and Influence Maps. In S. Rabin (Ed.), AI Game Programming Wisdom 2. Hingham, MA: Charles River Media, Inc., pp. 439-446.

Sweetser, P. & Wiles, J. (unpublished) Using Cellular Automata to Facilitate Emergence in Game Environments. Submitted to the Journal of Game Development.

Tozour, P. (2001) Influence Mapping. In M. Deloura (Ed.), Game Programming Gems 2. Hingham, MA: Charles River Media, Inc., pp. 287-297.