evolving aesthetic maps for a real time strategy game
DESCRIPTION
This study presents a procedural content generator method that have been able to generate aesthetic maps for a real-time strategy game. The maps has been characterized based on several of their properties in order to define a similarity function between scenarios. This function has guided a multi-objective evolution strategy during the process of generating and evolving scenarios that are similar to other aesthetic maps while being different to a set of non-aesthetic scenarios. The solutions have been checked using a support-vector machine classifier and a self-organizing map obtaining successful results (generated maps have been classified as aesthetic maps).TRANSCRIPT
Evolving Aesthetic
Maps for a Real-Time
Strategy Game
Raúl Lara-Cabrera, Carlos Cotta and Antonio J. Fernández-Leiva
Dpto. Lenguajes y ciencias de la computación
Universidad de Málaga
Procedural Content
Generation (PCG)
Automated production of game content by pseudo-random
process
Maps and levels Weapons, ítems, … Music and effects
Procedural Content
Generation (PCG)
The game: Planet Wars
Automatic map generator
EVOLUTIONARY ALGORITHM
Self-adaptive evolutionary
strategy (ES)
(µ+λ) generational scheme
with µ=10 y λ=100
Binary tournament selection
Individual’s evaluation is
based on game statistics
gathered by the
tournament system
TOURNAMENT SYSTEM
It executes games that take
place on generated maps
and gathers statistics
3 artificial players (bots)
ranked in the AI Challenge
Top 10 (Planet Wars)
Evolutionary algorithm:
representation
Evolutionary algorithm:
operators
Cut & splice recombination:
Gaussian mutation (for real-valued parameters) and
geometric mutation (for integer parameters):
Parents
Children
Improving maps’ aesthetics
Although we get balanced and dynamic maps, they have
bad aesthetics
The idea: use an evaluation function that measures how
aesthetic a map is
Measuring aesthetics
Evaluation and results
Two sets of maps labeled by
an expert:
Aesthetic
Non-aesthetic
Evolutionary multi-objective
algorithm:
Minimize the euclidean
distance to the aesthetic
maps set
Maximize the euclidean
distance to the non-
aesthetic maps set
Validation
SVM
Aesthetic/non-aesthetic
clasiffier
Training set: maps included
in aesthetic and non-
aesthetic sets
Every non-dominated
solution (4289) is classified
as: aesthetic
SELF-ORGANIZED NETWORK
Some examples
Conclusions
Initial approach towards the procedural aesthetic map
generation
We have defined a method of map characterization based on
several of its maps’ geometric and morphologic properties
Two set of maps (aesthetics and non-aesthetics) as a baseline
to compare generated maps with
Evolution strategy whose objectives are minimize and
maximize the distance of the generated maps to the aesthetic
and non-aesthetics maps of the baseline
The solutions have been tested with a SVM (solutions classified
as aesthetic) and a SOM (solutions share the same region as
aesthetic-maps)
ANYSELFTIN2011-28627-C04-01(Spanish MICINN)
DNEMESISTIC-6083 (Junta deAndalucía)